Cascade Transform fusion method and system based on multi-source heterogeneous geological data

Abstract

The invention relates to the field of geological data, in particular to a cascade Transform fusion method and system based on multi-source heterogeneous geological data. The method comprises the steps of obtaining a historical paper geological data set and a real-time multi-source geological monitoring data stream, and generating a multi-source heterogeneous geological feature set based on the historical paper geological data set and the real-time multi-source geological monitoring data stream; based on this, performing coarse-to-fine multi-level feature extraction and cross-modal attention fusion through a cascaded Transform network, and generating a deep fusion feature field; on the basis, multi-target collaborative optimization calculation is carried out by integrating geological prior knowledge rules, and a risk-resource integrated three-dimensional geological model for quantifying disaster risk probability and resource economic value is generated; on the basis, concealed disaster targeted treatment and resource re-exploitation optimization are carried out through a decision mapping engine, and a collaborative optimization engineering scheme set is generated. In the data cascade fusion process, the intelligent level and comprehensive benefits of geological engineering activities are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of geological data, in particular to a cascading Transformer fusion method and system based on multi-source heterogeneous geological data. BACKGROUND

[0002] In the field of deep mineral resource exploration and mine safety collaborative management, positioning the remaining ore and disaster risk warning in the existing deep and peripheral mines is the core technology to activate resource stock, ensure production safety and extend the service life of the mine, and has become a key link to promote the sustainable development of the mining industry.

[0003] However, the existing geological data analysis and modeling methods lack a mechanism for unified representation, dynamic fusion and collaborative interpretation of deep temporal and spatial correlations and physical coupling relationships between data when faced with multi-source, heterogeneous, and spatiotemporal non-uniform geological data. They not only cannot effectively support the accurate delineation of concealed resource occurrence locations, but also cannot quantitatively assess potential disaster risks. SUMMARY

[0004] The present application provides a cascading Transformer fusion method and system based on multi-source heterogeneous geological data to solve the above technical problems.

[0005] In a first aspect, the present application provides a cascading Transformer fusion method based on multi-source heterogeneous geological data, which comprises: Obtaining a set of historical paper geological data and a real-time multi-source geological monitoring data stream, structurally analyzing and spatiotemporally encoding the set of historical paper geological data, and extracting features and sequencing the real-time multi-source geological monitoring data stream to generate a set of multi-source heterogeneous geological features; Based on the set of multi-source heterogeneous geological features, performing multi-level feature extraction and cross-modal attention fusion from coarse to fine through a cascading Transformer network to generate a deep-level fusion feature field; Based on the deep-level fusion feature field, integrating geological priori knowledge rules for multi-objective collaborative optimization calculation to generate a risk-resource integrated three-dimensional geological model of quantitative disaster risk probability and resource economic value; Based on the risk-resource integrated three-dimensional geological model, performing optimization of concealed disaster targeted management and resource re-mining through a decision mapping engine to generate a set of collaborative optimization engineering schemes.

[0006] Through the above technical solution, based on automatic and intelligent scheme generation, the efficiency and scientificity of engineering planning are significantly improved, and deep collaboration of disaster management and resource development in space, time sequence and benefit is realized, providing a powerful intelligent decision-making tool for mine safety production, urban underground space safety utilization, and major project site selection.

[0007] Optionally, the generating the multi-source heterogeneous geological feature set comprises: performing structured analysis and coding on the historical paper geological data set based on optical character recognition, visual model and spatio-temporal correlation to generate a structured mine historical knowledge graph; performing multi-modal feature extraction and spatio-temporal alignment processing on the real-time multi-source geological monitoring data stream to generate a multi-source observation spatio-temporal feature tensor of a unified reference; based on the entity attribute feature vector extracted from the mine historical knowledge graph and the multi-source observation spatio-temporal feature tensor, generating the multi-source heterogeneous geological feature set.

[0008] Optionally, the generating the structured mine historical knowledge graph comprises: the historical paper geological data set includes old mine geological exploration report text, old mine hand-drawn geological section map, old mine drilling log table and historical mining record table; extracting standardized geological entities and geological semantic relationships between the geological entities from the old mine geological exploration report text through a cascade processing of an optical character recognition model and geological field professional terms; identifying line symbols in the old mine hand-drawn geological section map through a visual transformer, and analyzing and reconstructing them into vectorized graphic objects with spatial coordinates and geological attributes; aligning the text structured data in the old mine drilling log table and the historical mining record table with the entities extracted from the old mine geological exploration report text and the old mine hand-drawn geological section map in space and time; based on the geological semantic relationships, the vectorized graphic objects and the aligned text structured data, constructing the mine historical knowledge graph.

[0009] Optionally, the generating the multi-source observation spatio-temporal feature tensor of a unified reference comprises: the real-time multi-source geological monitoring data stream includes microseismic monitoring waveform time series data, geological chemical element sampling point data and engineering drilling core logging data; extracting source parameters from the microseismic monitoring waveform time series data based on an event detection algorithm to form a microseismic event feature sequence in units of events; processing the geological chemical element sampling point data based on spatial interpolation and outlier analysis algorithm to generate an element concentration spatial distribution feature map; analyzing the engineering drilling core logging data based on natural language processing and lithology classification model to generate a lithology-structure discrete feature sequence with depth coordinates; unify the microseismic event feature sequence, the element concentration spatial distribution feature map and the lithology-structure discrete feature sequence into a same three-dimensional geographic coordinate system and time reference to generate the multi-source observation spatio-temporal feature tensor.

[0010] Optionally, the generating the deep-level fusion feature field comprises: inputting the multi-source heterogeneous geological feature set into a plurality of independent and structure-specialized cascaded Transformer networks, the cascaded Transformer networks comprising single-source feature extraction layers, shallow cross-modal interaction layers and deep cross-modal fusion layers connected in sequence; in the single-source feature extraction layer, performing preliminary abstract feature analysis on the entity attribute feature vector from the mine historical knowledge graph and different modal feature sequences from the multi-source observation spatio-temporal feature tensor by independent Transformer encoders; in the shallow cross-modal interaction layer, introducing a cross-modal attention mechanism to perform low-dimensional interaction and alignment on the preliminarily abstracted modal features; in the deep cross-modal fusion layer, performing high-dimensional semantic feature reorganization and deep fusion through a multi-head cross-attention mechanism to output the deep-level fusion feature field representing the geological spatio-temporal structure and attributes.

[0011] Optionally, the single-source feature extraction layer comprises: deploying Transformer encoders with different induction biases for different modal data directly related to ore-forming potential to perform the preliminary abstract feature screening for ore formation: using a graph structure enhanced Transformer encoder to process the entity attribute feature vector to strengthen the encoding of topological relationships among historical ore bodies, ore-controlling structures and ore-forming favorable lithology entities, to obtain historical geological entity features; using a time sequence perception Transformer encoder to process the microseismic event feature sequence to extract its spatio-temporal aggregation and energy release trend features, to obtain microseismic activity features; using a space perception Transformer encoder to process the element concentration spatial distribution feature map to identify spatial morphology and concentration gradient features of element combination anomalies, to obtain geochemical anomaly features.

[0012] Optionally, the shallow cross-modal interaction layer comprises: introducing a lightweight cross-modal attention module with spatial position as the key to perform initial cross-validation and positioning focus of prospecting clues: The preliminary abstract features of each modality are projected into a unified gridded space based on their three-dimensional spatial coordinates, and the mutual attention weights between different modality features are analyzed through a lightweight cross-attention mechanism; Based on the mutual attention weights, the historical geological entity features, the microseismic activity features and the geochemical anomaly features are weightedly fused and information-complemented to generate spatially preliminarily coupled anomaly-structure joint features for delineating ore-forming favorable areas.

[0013] Optionally, the deep cross-modal fusion layer comprises: Taking the spatially coupled anomaly-structure joint features as input, a mineralization mode feature vector extracted from the mine historical knowledge graph is introduced as a global query; The deep semantic matching degree of the mineralization mode feature vector and each spatial position joint is analyzed through a multi-head cross-attention mechanism; The deep semantic matching degree is taken as a fusion feature vector of each spatial position in the deep fusion feature field, and the fusion feature vector directly represents the similarity between the spatial position and the historical mineralization mode, i.e., the potential probability of the remaining ore occurrence.

[0014] Optionally, the risk-resource integrated three-dimensional geological model for generating quantitative disaster risk probability and resource economic value comprises: Based on the deep fusion feature field, a dual-path feature representation for disaster risk prediction and resource potential evaluation is extracted in parallel; In the disaster risk quantification path, the deep fusion feature field is constraintly fused with the geomechanics prior rules, and the disaster risk probability value is generated voxel by voxel through the coupling calculation of the concealed structure activation degree and the rock mass stability parameters; In the resource economic value quantification path, the fusion feature vector representing the mineralization mode matching degree is coupled with the ore grade, the mining cost and the market parameters for multi-objective optimization, and the resource economic value index is generated voxel by voxel; The disaster risk probability value and the resource economic value index are voxel-level integrated under a unified three-dimensional geological space framework to generate a risk-resource integrated three-dimensional geological model containing both risk and value quantification attributes for each spatial position.

[0015] In a second aspect, the present application provides a cascading Transformer fusion system based on multi-source heterogeneous geological data, which comprises: A geological feature analysis module is configured to acquire a historical paper geological data set and a real-time multi-source geological monitoring data stream, to perform structural analysis and spatio-temporal coding on the historical paper geological data set, and to perform feature extraction and serialization on the real-time multi-source geological monitoring data stream to generate a multi-source heterogeneous geological feature set. a feature field generation module configured to generate a deep-level fusion feature field by performing coarse-to-fine multi-level feature extraction and cross-modal attention fusion through a cascaded Transformer network based on the multi-source heterogeneous geological feature set; a three-dimensional module construction module configured to generate a risk-resource integrated three-dimensional geological model quantifying disaster risk probability and resource economic value by performing multi-objective collaborative optimization calculation based on the deep-level fusion feature field and incorporating geological priori knowledge rules; an optimization scheme generation module configured to generate a set of collaborative optimization engineering schemes by performing optimization of hidden disaster targeted management and resource re-mining through a decision mapping engine based on the risk-resource integrated three-dimensional geological model. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0017] Figure 1 An application scenario schematic diagram provided by an embodiment of the present application; Figure 2 A flowchart of a cascaded Transformer fusion method based on multi-source heterogeneous geological data provided by an embodiment of the present application; Figure 3 A structural schematic diagram of a cascaded Transformer fusion system based on multi-source heterogeneous geological data provided by an embodiment of the present application. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0019] In addition, the term "and / or" in this paper is only to describe the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper, unless otherwise specified, generally represents an "or" relationship between the associated objects before and after it.

[0020] The embodiments of the present application are further described in detail below with reference to the accompanying drawings.

[0021] The existing geological data analysis and modeling method lacks a mechanism for unified representation, dynamic fusion and collaborative interpretation of deep-level spatio-temporal correlation and physical coupling relationship among data when facing multi-source, heterogeneous, spatio-temporal non-uniform geological data, which cannot effectively support the accurate delineation of the location of concealed resources, and cannot quantitatively evaluate potential disaster risks.

[0022] Based on this, the present application provides a cascading Transformer fusion method and system based on multi-source heterogeneous geological data. First, historical paper materials and real-time monitoring data streams are accessed in parallel, and NLP and signal processing techniques are used for structured analysis, spatio-temporal coding and feature serialization, to construct a multi-source heterogeneous geological feature set under a unified spatio-temporal framework. Then, through a cascading Transformer network, multi-level cross-modal attention calculation is performed to realize data-driven deep feature fusion from local details to global rules, and generate a deep-level fusion feature field. Further, formalized geological prior knowledge rules are used as constraints, and a multi-objective collaborative optimization algorithm is used to generate an integrated three-dimensional geological model that quantitatively evaluates disaster risks and resource values. Finally, based on the model, the decision mapping engine uses intelligent optimization algorithms to automatically search and generate a Pareto optimal engineering solution set that coordinates hidden disaster targeted governance and resource re-exploitation, and outputs it to geological researchers. This method solves the long-standing data island and information sleep problems in the geological field, generates a quantitative three-dimensional model that integrates risks and resources, provides a unified scientific decision-making base for precise early warning and prevention and control of hidden disasters and safe and green development of mineral resources, and significantly improves the intelligent level and comprehensive benefits of geological engineering activities.

[0023] Figure 1 An application scenario diagram is provided for the present application. In the data cascading fusion process, the method provided by the present application provides a scientific decision-making base for the safe and green development of mineral resources, and significantly improves the intelligent level and comprehensive benefits of geological engineering activities.

[0024] In particular, the method of the present application is applied to any server in communication with a mine history archive and a geological monitoring sensor network, through which a set of historical paper geological data provided by the mine history archive and a real-time multi-source geological monitoring data stream provided by the geological monitoring sensor network are obtained. First, the historical paper data and the real-time monitoring data stream are accessed in parallel, and NLP and signal processing techniques are used for structured analysis, spatio-temporal coding and feature serialization, to construct a multi-source heterogeneous geological feature set under a unified spatio-temporal framework. Then, through a cascaded Transformer network, multi-level cross-modal attention calculation is performed to realize data-driven deep feature fusion from local details to global rules, generating a deep-level fusion feature field. Further, formalized geological prior knowledge rules are used as constraints, and a multi-objective collaborative optimization algorithm is used to generate an integrated three-dimensional geological model that quantifies disaster risk and resource value. Finally, based on the model, the decision mapping engine uses intelligent optimization algorithms to automatically search and generate a set of Pareto optimal engineering solutions that coordinate hidden disaster targeted management and resource re-mining, and outputs them to geological researchers.

[0025] The specific implementation can refer to the following embodiments.

[0026] Figure 2 The flowchart of the cascaded Transformer fusion method based on multi-source heterogeneous geological data provided by an embodiment of the present application. The method of the present embodiment can be applied to the server in the above scenario. As shown in the figure, the method comprises: Figure 2 S201, obtaining a set of historical paper geological data and a real-time multi-source geological monitoring data stream, and performing structured analysis and spatio-temporal coding on the set of historical paper geological data, and performing feature extraction and serialization on the real-time multi-source geological monitoring data stream, to generate a multi-source heterogeneous geological feature set.

[0027] The set of historical paper geological data can be a collection of various types of geological-related data accumulated over a long period of time in the form of paper carriers, including old mine geological exploration report texts, old mine hand-drawn geological profile maps, old mine drilling log tables and historical mining records, and the data is sourced from the mine history archive retained by the enterprise. The real-time multi-source geological monitoring data stream can be a continuous transmission stream of geological environment and geological state related data from different monitoring dimensions collected in real time by various types of geological monitoring equipment, including microseismic monitoring waveform time series data, geological chemical element sampling point data and engineering drilling core logging data, and the data is sourced from various types of geological monitoring sensor networks deployed in the monitoring area. The multi-source heterogeneous geological feature set can be a set of historical geological data features integrated through structured analysis and spatio-temporal coding, and real-time geological monitoring data features integrated through feature extraction and serialization.

[0028] ​Specifically, in the fields of geological exploration, disaster prevention and resource development, the existing technology has long been faced with the prominent problems of insufficient utilization of data sources and poor data fusion, which seriously restricts the comprehensiveness and accuracy of geological analysis: On the one hand, historical paper geological data is a valuable treasure accumulated by long-term geological research and engineering practice, which contains non-renewable geological basic information. However, due to its carrier being paper, it has natural defects such as storage dispersion, retrieval difficulty and inability to be directly processed by computers, which leads to the long-term "sleeping" state of these historical data, making it difficult to form effective complementation with real-time monitoring data, resulting in serious waste of geological data resources. For example, the paper survey archives retained by a certain mining enterprise for decades contain key information such as early stratum distribution and structural characteristics in the region, but due to the inability to quickly analyze and utilize, they are not fully referenced in subsequent resource exploitation planning, resulting in deviations between the exploitation plan and the actual geological conditions, increasing the exploitation risk and cost. On the other hand, with the rapid development of geological monitoring technology, various monitoring equipment has emerged, forming a multi-source heterogeneous real-time data stream. These data are different in source, format and dimension. Traditional data processing methods can only process simple data of a single type or a single source, lacking effective mining of associated features in multi-source data, making it difficult to fully release the value of data. This step collects historical paper geological data and real-time multi-source monitoring data stream. The paper data is scanned, OCR extracted and structured, and then time and space coding is added. The real-time data is preprocessed to extract core features and serialized. Finally, the two types of features are integrated to generate a multi-source heterogeneous geological feature set. The standardization, structuring and spatio-temporal integration of historical geological data and real-time monitoring data convert the sleeping paper archives and experiential description into computable data, and refine the high-frequency, multi-dimensional monitoring data into feature sequences with clear geological significance, providing high-quality, semantically associated inputs for subsequent deep information fusion, fundamentally changing the situation of dispersed, heterogeneous and difficult-to-utilize geological data.

[0029] S202, based on the multi-source heterogeneous geological feature set, a coarse-to-fine multi-level feature extraction and cross-modal attention fusion is performed through a cascaded Transformer network to generate a deep-level fusion feature field.

[0030] The cascaded Transformer network can be a deep learning network model composed of multiple Transformer sub-networks in a hierarchical progressive manner, each sub-network undertaking feature extraction tasks at different levels, and the output of the previous sub-network is used as the input of the next sub-network. Multi-level feature extraction can be a feature extraction process that unfolds according to the logic of from the whole to the local and from the surface to the core. First, the primary Transformer sub-network extracts shallow and general features (coarse-grained features) in the multi-source and heterogeneous geological feature set. Then, the subsequent Transformer sub-networks gradually mine the details, associated information and potential laws (fine-grained features) in the features, achieving in-depth feature extraction. Cross-modal attention fusion can be the calculation of the correlation weights between different modalities (such as historical structured data, real-time sensor data, image data, etc.) in the multi-source and heterogeneous geological feature set through the attention mechanism. Deeply fused feature field can be a high-dimensional and integrated feature space that can comprehensively and deeply reflect the geological historical background, real-time dynamic changes and internal correlation laws of each geological factor in the monitoring area after multi-level extraction and cross-modal fusion by the cascaded Transformer network, containing features at all levels from coarse-grained to fine-grained and associated information between modalities.

[0031] Specifically, the existing geological feature fusion technology generally has the problems of insufficient fusion depth and poor modality adaptability, which cannot meet the needs of complex geological system analysis: on the one hand, traditional feature extraction methods (such as convolutional neural networks, single Transformer network, etc.) can usually only extract single-level or single-modal geological features, and cannot achieve step-by-step mining from shallow to deep, resulting in that the extracted features are difficult to fully reflect the complex nature of geological phenomena. For example, the surface stratum features extracted by a single network cannot reveal the potential impact of deep geological structures on disaster occurrence, making the subsequent disaster risk assessment lack key basis; on the other hand, multi-source heterogeneous geological data contains multiple different modalities, and the structure, dimension, and semantic information of different modal data are significantly different. Traditional fusion methods (such as simple concatenation, weighted summation, etc.) lack effective mining of associated information between modalities, and are prone to problems such as modal feature conflict and effective information being overwhelmed, resulting in low quality of the fused features, which in turn affects the accuracy of subsequent geological model construction. For example, simple concatenation of historical structured stratum data and real-time image monitoring data cannot reflect the association between the two in the time and space dimensions, and the fusion effect is greatly discounted. This step takes a multi-source heterogeneous geological feature set as input, extracts coarse-grained shallow features through a cascaded Transformer with multiple sub-networks in series, and then gradually mines fine-grained deep associations. Each level embeds cross-modality attention fusion to generate a deep-level fusion feature field reflecting the complex characteristics of geology. Using the powerful modeling capability of the cascaded Transformer, the multi-source heterogeneous geological data is bridged from shallow association to deep semantic fusion, providing a highly information-dense and highly complex-association unified feature representation for subsequent quantitative calculation and decision-making.

[0032] S203, based on the deep-level fusion feature field, multi-objective collaborative optimization calculation is performed by incorporating geological priori knowledge rules to generate a risk-resource integrated three-dimensional geological model of quantitative disaster risk probability and resource economic value.

[0033] The geological priori knowledge rules can be a set of deterministic or statistical rules formed based on long-term geological research, engineering practice experience and industry standard specifications, which can reflect the evolution law of geological phenomena, the interaction relationship of geological factors, the disaster occurrence mechanism and the distribution law of resources. The risk-resource integrated three-dimensional geological model can be constructed based on the deep-level fusion feature field and the geological priori knowledge rules through multi-objective collaborative optimization calculation, which can intuitively present an integrated digital model of the geological structure, stratum distribution, disaster risk probability distribution and resource economic value distribution in the monitoring area in three-dimensional space.

[0034] Specifically, in traditional geological model construction, there are prominent problems of "single target orientation" and "lack of prior knowledge guidance", which seriously affect the practicability and reliability of the model, and it is difficult to meet the core needs of modern geological engineering "safety and efficiency". Existing geological models often only focus on disaster risk assessment or resource value assessment, and it is difficult to realize the simultaneous quantification and integrated presentation of the two. For example, the traditional disaster risk assessment model only focuses on the occurrence probability and influence range of geological disasters, but ignores the resource distribution and economic value of the region, leading to excessive investment in cost when formulating disaster management schemes, and excessive management measures for high-value resource areas, resulting in resource waste; while the traditional resource assessment model focuses on the reserves and economic value of resources, ignoring the geological disaster risks that may be faced in the mining process, lacking targeted risk prevention and control considerations, and easily causing safety accidents, causing personnel casualties and significant economic losses. This step builds a geological prior knowledge rule base, extracts core variables from the deep fusion feature field, establishes a multi-objective optimization function with rule constraints, solves the quantitative disaster risk and resource value, and combines three-dimensional modeling technology to generate a risk-resource integrated three-dimensional geological model. The combination of geological prior knowledge and data-driven models gives artificial intelligence "geological common sense", and through multi-objective collaborative optimization, a three-dimensional integrated model that simultaneously quantifies and evaluates safety risks and economic values is generated. This model breaks down the barriers between traditional geological modeling, risk assessment, and resource evaluation, providing a scientific and comprehensive digital foundation that can directly serve comprehensive decision-making.

[0035] S204, based on the risk-resource integrated three-dimensional geological model, the decision mapping engine is used to optimize the hidden disaster targeted management and resource re-mining, and a set of collaborative optimization engineering schemes is generated.

[0036] The set of collaborative optimization engineering schemes can be a set containing multiple collaborative optimization schemes for hidden disaster targeted management and resource re-mining.

[0037] Specifically, the traditional engineering scheme formulation process lacks intelligent and visual decision support tools. The scheme is formulated as a static design, which cannot be adjusted in real time according to the dynamic changes of the geological model, and it is also difficult to simulate the implementation effect of different schemes, resulting in poor adaptability of the formulated scheme, limited optimization space, and difficulty in meeting the dynamic needs of complex geological engineering. In this step, the three-dimensional model data is input into the decision mapping engine, the high-risk and high-value areas are identified and analyzed, and the implementation details of each scheme are optimized after screening out non-compliant schemes to form a set of collaborative optimization engineering schemes. The advanced geological cognition model is directly converted into an executable engineering blueprint, realizing the full-chain closed loop from “data -> model -> knowledge -> decision”. Through automatic and intelligent scheme generation, the efficiency and scientificity of engineering planning are significantly improved, realizing deep collaboration in space, time sequence and benefit between disaster control and resource development, and providing a powerful intelligent decision-making tool for mine safety production, safe use of urban underground space, and major project site selection.

[0038] By the way provided by the embodiment, firstly, historical paper materials and real-time monitoring data streams are accessed in parallel, and NLP and signal processing technologies are used for structured analysis, spatio-temporal coding and feature serialization, to construct a multi-source heterogeneous geological feature set under a unified spatio-temporal framework; then through a cascaded Transformer network, multi-level cross-modal attention calculation is performed to realize data-driven deep feature fusion from local details to global rules, and generate a deep-level fusion feature field; further, formalized geological prior knowledge rules are used as constraints, and a multi-objective collaborative optimization algorithm is used to generate an integrated three-dimensional geological model that quantifies disaster risk and resource value; finally, based on the model, the decision mapping engine uses an intelligent optimization algorithm to automatically search and generate a Pareto optimal engineering scheme set that coordinates hidden disaster targeted control and resource re-mining, and outputs it to geological researchers. This method solves the long-standing problems of data silos and information dormancy in the geological field, generates a quantitative three-dimensional model integrating risk and resources, provides a unified scientific decision-making base for precise early warning and prevention and control of hidden disasters and safe and green development of mineral resources, and significantly improves the intelligent level and comprehensive benefits of geological engineering activities.

[0039] In some embodiments, the historical paper geological material set is structured and analyzed based on optical character recognition, visual models and spatio-temporal correlation, to generate a structured mine history knowledge graph; the real-time multi-source geological monitoring data stream is processed for multi-modal feature extraction and spatio-temporal alignment, to generate a unified reference multi-source observation spatio-temporal feature tensor; based on the entity attribute feature vector extracted from the mine history knowledge graph and the multi-source observation spatio-temporal feature tensor, a multi-source heterogeneous geological feature set is generated.

[0040] Optical character recognition can be a technology that converts printed or handwritten text in an image to machine-encodable text. A vision model can be a deep learning-based computer vision model, such as an image segmentation model, an object detection model, etc. Structured parsing and encoding can be a process of converting unstructured raw materials (text, drawings) into structured data with explicit semantic relationships and spatiotemporal labels through OCR, vision models, and other technical means. A mine history knowledge graph can be a knowledge network that presents mine historical geological information in a structured form, including entities in the field of mine geology (such as strata, structures, mineral resource points, etc.), entity attributes (such as stratum thickness, lithology type, resource reserves, etc.), and the correlation between entities (such as the spatial position relationship between strata and structures, the attribution relationship between resource points and strata, etc.), which is the core achievement of structured parsing and encoding of historical paper geological data. Multi-modal feature extraction can be a processing technology for different types of data (modalities) in real-time multi-source geological monitoring data streams. Spatiotemporal alignment processing can be due to the different sampling frequencies, deployment locations, coordinate systems, and data formats of different monitoring devices (sources), and the spatiotemporal alignment processing aims to unify these multi-source, heterogeneous real-time monitoring data to the same spatiotemporal reference. Multi-source observation spatiotemporal feature tensor can be after multi-modal feature extraction and spatiotemporal alignment processing, all monitoring features under the same spatiotemporal unit (such as a certain time slice, a certain three-dimensional geological unit) are organized to form a multidimensional data array (tensor).

[0041] Specifically, traditional mine geological analysis has long been plagued by data fragmentation: massive amounts of historical paper materials (such as hand-drawn cross-section maps, drilling logs) are non-structured and lie dormant in archives, forming a "data tomb" that cannot be directly understood by machines, and the ore-forming rules and disaster prior knowledge contained therein are difficult to reuse; while real-time monitoring data streams (such as microseisms, deformations) are dynamic and rich, but due to multi-source heterogeneity and different time and space references, they lack association with historical background, resulting in analysis that stops at isolated alarms and surface associations, and cannot reveal the causal mechanisms between "current anomalies" and "historical structures". To address the above problems, this step first collects mine historical paper materials (such as survey reports, drilling drawings), which are scanned and repaired in high definition, and then uses optical character recognition (OCR) technology to extract text values (such as "drilling depth 150m, stratum lithology is sandstone"), uses a convolutional neural network (visual model) to analyze the drawings and extract fault spatial coordinates, and then adds time and space codes according to a unified standard (time stamp set to UTC 2024-06-01T09:00:00, spatial coordinates use the national 2000 geodetic coordinate system) to generate a mine historical knowledge graph, from which stratum entity attribute feature vectors (such as "thickness 8m, compressive strength 30MPa") are extracted; at the same time, real-time monitoring data (such as displacement, audio) are received, a unified reference is set (time is unified to UTC, displacement unit is unified to millimeters, spatial coordinates are the same as historical data), the sliding window method is used to extract the average displacement data (such as 0.3mm / h), CNN is used to extract the crack features of the drilling wall image (such as width 2mm), Fourier transform is used to extract the rock vibration audio frequency spectrum peak value (such as 50Hz), and then 1-minute interval data is completed by linear interpolation, and coordinates are calibrated by spatial interpolation (such as mapping the 2-meter deviated device data to the grid node), generating a multi-source observation space-time feature tensor; finally, through a space-time association algorithm, the stratum entity vector is associated and fused with the corresponding real-time soil copper content data (such as 25mg / kg) in the region, and after removing redundant features, a multi-source heterogeneous geological feature set is generated.

[0042] In the manner provided by this embodiment, through the association and fusion of entity attribute feature vectors and multi-source observation space-time feature tensors, a multi-source heterogeneous geological feature set is generated, realizing the deep association and complementarity of historical geological data and real-time monitoring data, so that the feature set contains both historical geological evolution rules and current geological dynamic changes, fully reflecting the complex characteristics of the mine geological system.

[0043] In some embodiments, the historical paper geological data set includes old mine geological exploration report texts, old mine hand-drawn geological section maps, old mine drilling log tables, and historical mining record tables; through cascade processing of an optical character recognition model and professional terms in the geological field, standardized geological entities and geological semantic relationships between the geological entities are extracted from the old mine geological exploration report texts; through a visual transformer, line symbols in the old mine hand-drawn geological section maps are identified, and are parsed and reconstructed into vectorized graphic objects with spatial coordinates and geological attributes; text structured data in the old mine drilling log tables and the historical mining record tables are spatiotemporally associated and aligned with the entities extracted from the old mine geological exploration report texts and the old mine hand-drawn geological section maps; based on the geological semantic relationships, the vectorized graphic objects, and the aligned text structured data, a mine historical knowledge graph is constructed.

[0044] The old mine geological exploration report texts can be written report files formed after geological survey work is carried out in the old mine area. The old mine hand-drawn geological section maps can be graphical data formed by field surveying and mapping by historical geological workers, reflecting spatial relationships such as vertical or horizontal direction geological structure, stratum distribution, and ore body position of the old mine. The old mine drilling log tables can be tabularized paper records formed by recording immediate observation and description of the drilled core by field technical personnel during drilling construction. The historical mining record tables can be ledgers or reports recording information such as stope position, mining amount, ore grade, and encountered geological problems (such as small faults, water points) in past production activities of the mine. The optical character recognition model can be a computer vision model based on deep learning, used to detect text regions in scanned document images and identify them as editable and processable coded text. The professional terms in the geological field can be a series of fine text processing procedures specially performed for the geological field after general OCR recognizes the text. The geological semantic relationship can be a logical association with a clear geological meaning connecting different geological entities, which is an edge connecting nodes in the knowledge graph. The visual transformer can be a deep learning model that applies the successful Transformer architecture in natural language processing to image recognition tasks. The line symbol can be a graphical mark representing different geological elements in the old mine hand-drawn geological section map, such as solid lines representing certain stratum boundaries, dashed lines representing presumed boundaries, sawtooth lines representing faults, and various points, circles, and filled patterns representing different lithology. The vectorized graphic object can be a conversion of the identified line symbol from a raster image format to a vector graphic format (such as SVG path) defined by mathematical equations.

[0045] Specifically, in traditional mine geological analysis, historical paper materials (such as exploration report texts, hand-drawn drawings, and logging records) are in a fragmented state, forming multiple information islands that cannot be interconnected: the geological description in the text, the line symbols on the drawing, and the specific data in the table are isolated from each other, relying on the manual memory, review, and subjective association of geologists for comprehensive research and judgment, which is extremely inefficient and prone to key cognitive gaps due to personnel changes or missing details. Simply scanning and archiving the data only achieves digitization, far from knowledge. The machine cannot understand that "F5 fault mentioned in the report" and "a jagged line on the profile drawing" and "angular conglomerate at a depth of 120 meters in the logging table" are the same entity. To address the above problems, this step first collects old mine geological exploration report texts, hand-drawn geological profiles, drilling logging tables, and historical mining records, and after repair, converts them into electronic images using a high-definition scanner (resolution set to 300 dpi) and enhances them; input the report text into an optical character recognition model to extract the original text, cascade processing with a geological term library (including "ore-bearing layer" "conformable contact" and other words), correct errors such as "sandstone" to "sandstone", extract entities such as "F2 normal fault" "Jurassic sandstone formation" and semantic relationships such as "F2 normal fault controls the western ore body"; use a visual transformer to analyze the profile drawing, identify the 1:5000 scale and fault line symbols, and reconstruct them as vectorized objects with coordinates; extract the structured data from the drilling log and mining records, associate them with the 2000 mine area coordinate system and the 1988 time base, and finally use a graph database to build a knowledge graph, with nodes containing entity coordinates and other attributes, and edges representing semantic relationships.

[0046] In the manner provided by this embodiment, by deeply associating text, drawing, and table data, this embodiment forms a unified and authoritative digital twin of the mine geological conditions, enabling subsequent risk assessment, resource evaluation, and engineering planning to be based on complete and consistent historical knowledge, significantly improving the value of old mine data reuse and the level of intelligent management in the new era.

[0047] In some embodiments, the real-time multi-source geological monitoring data stream includes microseismic monitoring waveform time series data, geological chemical element sampling point data, and engineering drilling core logging data; event detection algorithms are used to extract source parameters from the microseismic monitoring waveform time series data to form a microseismic event feature sequence based on events; spatial interpolation and outlier analysis algorithms are used to process the geological chemical element sampling point data to generate an element concentration spatial distribution feature map; natural language processing and lithology classification models are used to analyze the engineering drilling core logging data to generate a lithology-structure discrete feature sequence with depth coordinates; the microseismic event feature sequence, the element concentration spatial distribution feature map, and the lithology-structure discrete feature sequence are converted to the same three-dimensional geographic coordinate system and time base, generating a multi-source observation spatiotemporal feature tensor.

[0048] The microseismic monitoring waveform time series data can be waveform signals generated by microseismic sensors in real time, and the waveform signals are data of small vibrations caused by rock mass rupture, tectonic activity, etc. in the mining area. The geological and chemical element sampling point data can be data obtained by laboratory testing or on-site rapid testing after collecting rock, soil or groundwater samples at different spatial positions in the mining area, representing the content of various chemical elements (such as copper, lead, zinc, sulfur, etc. related to mineral resources or geological environment) in the sample. The engineering drilling core logging data can be data formed by on-site observation, description and recording of the core obtained by drilling during the construction process of the engineering drilling in the mining area, including core depth, lithological characteristics, tectonic phenomena, mineral development conditions, etc. Text information is the core data reflecting the true state of underground rock layers. The microseismic event feature sequence can be an ordered feature set formed by sorting a plurality of microseismic events extracted by an event detection algorithm according to the time of occurrence, each event corresponding to a set of source parameters. The spatial interpolation and outlier analysis algorithm can be a combined algorithm composed of a spatial interpolation algorithm (such as Kriging interpolation, inverse distance weighted interpolation) and an outlier analysis algorithm (such as Grubbs test, box plot analysis). The element concentration spatial distribution feature map can be an image that visually presents the concentration distribution of a certain type or multiple types of chemical elements in the spatial range (plane or three-dimensional space) of the mining area in the form of a graph after processing by a spatial interpolation algorithm. The natural language processing and lithology classification model can be a joint model composed of a natural language processing (NLP) model (such as BERT, LSTM) and a lithology classification model (such as ResNet, support vector machine). The lithology-tectonic discrete feature sequence can be a discrete feature set of “lithology category + tectonic feature” arranged in order of drilling depth after analyzing the engineering drilling core logging data, each depth section corresponding to a set of discrete features. The three-dimensional geographic coordinate system can be a three-dimensional coordinate system for unified representation of the spatial position of the mining area, including longitude, latitude, and elevation (or depth) three dimensions, ensuring that various types of spatial data can be associated under the same spatial reference.

[0049] Specifically, the traditional processing of real-time multi-source geological monitoring data stream has obvious defects: the microseismic monitoring waveform time series data is only simply filtered, without using event detection algorithm to extract source parameters, and cannot quantify the depth and magnitude of microseismic events; the geological chemical element sampling point data is stored in discrete form without spatial interpolation and outlier analysis, which is difficult to present the element distribution rule and is easily disturbed by outliers; the engineering drilling core logging data relies on manual recording, the lithology description is not standard and is not associated with depth; more importantly, the time and space benchmarks of the three types of data are chaotic and cannot be integrated. To solve the above problems, real-time data is first collected: microseismic sensors are deployed at intervals of 500 m to obtain waveform data with a sampling frequency of 1000 Hz; rock samples are collected in a 200 m x 200 m grid, and the element concentration (such as copper 360 mg / kg, lead 80 mg / kg) is measured using a portable detector; when drilling, a terminal with a depth sensor is used to record core data (such as ZK203 drilling 320-325 m: gray-black rock containing pyrite). Then process respectively: microseismic data uses STA / LTA event detection algorithm (short window 5 s, long window 20 s) to extract source parameters (such as depth 520 m, magnitude ML1.2) to form event sequence; element data uses Grubbs test (significance 0.05) to remove outliers (such as zinc 1850 mg / kg), and kriging interpolation generates 10 m x 10 m concentration map; core data uses BERT model to extract information, and ResNet lithology classification model corrects "gray-black rock" to basalt, and forms feature sequence according to depth. Finally, unify the national 2000 coordinate system and UTC time (such as correcting the local time of microseismic 13:10 to UTC 14:10), and integrate into multi-source observation space-time feature tensor.

[0050] Through the way provided by the embodiment, by unifying the three-dimensional geographic coordinate system and the time benchmark, the time and space barriers of the three types of data are broken, and the generated multi-source observation space-time feature tensor can simultaneously carry the time and space correlation information of microseismic, elements, and cores, providing high-quality and highly correlated input data for subsequent cascading Transformer network for multi-level feature extraction and cross-modal fusion, significantly improving the accuracy of deep-level fusion feature field.

[0051] In some embodiments, a plurality of source heterogeneous geological feature sets are input into a plurality of independent and structure-specialized cascaded Transformer networks, which include single-source feature extraction layers, shallow cross-modal interaction layers, and deep cross-modal fusion layers connected in sequence; in the single-source feature extraction layers, independent Transformer encoders are used to respectively perform preliminary abstract feature analysis on entity attribute feature vectors from a mine history knowledge graph and different modal feature sequences from a plurality of source observation spatio-temporal feature tensors; in the shallow cross-modal interaction layers, cross-modal attention mechanisms are introduced to perform low-dimensional interaction and alignment on the preliminarily abstracted modal features; in the deep cross-modal fusion layers, multi-head cross-attention mechanisms are used to perform high-dimensional semantic feature reorganization and deep fusion, and output deep-level fusion feature fields representing geological spatio-temporal structures and attributes.

[0052] The single-source feature extraction layer can be the basic level of the cascaded Transformer network, and the core function is to independently perform preliminary abstraction and feature extraction on the original features of a single modality, avoiding early interference between different modalities and laying a foundation for subsequent cross-modal interaction. The shallow cross-modal interaction layer can be the intermediate level after the single-source feature extraction layer, and the core function is to realize preliminary interaction, alignment, and correlation mining of different modal features in a low-dimensional feature space, solving the problems of semantic differences and spatio-temporal misplacement of different modal features. The deep cross-modal fusion layer can be the core level of the cascaded Transformer network, located after the shallow interaction layer, and the core function is to perform deep reorganization, correlation strengthening, and integrated fusion of the preliminarily aligned cross-modal features in a high-dimensional semantic space, mining deep geological semantic correlations between modalities. The cross-modal attention mechanism can be an attention computing mechanism that can calculate the correlation weights between different modal features, focus on key correlation information, weaken irrelevant interference, and realize feature alignment and interaction between modalities, which is the core technology of the shallow cross-modal interaction layer. The multi-head cross-attention mechanism can be a mechanism that calculates cross-correlation weights between cross-modal features from different dimensions and perspectives through multiple parallel attention heads, realizes deep reorganization and fusion of high-dimensional semantic features, and is the core technology of the deep cross-modal fusion layer.

[0053] Specifically, traditional multi-source geological data fusion techniques often use simple feature splicing or single model unified processing, ignoring the inherent physical meaning differences, scale gaps and asymmetric interpretation relationships between different modal data (such as descriptive historical knowledge, point-like microseismic events, continuous field element concentration, and sequential lithology profile). This leads to a mechanical mixing of articles in different languages in the fusion process, and the features cannot effectively communicate and constrain each other in semantics, for example, historical fault attributes cannot be used as prior background to guide and explain the spatio-temporal clustering pattern of microseismic events. As a result, the fused feature representation is mixed and poorly interpretable, and cannot support deep and reliable joint inversion of complex geological spatio-temporal structures and attributes. To solve the above problems, the step first acquires a set of multi-source heterogeneous geological features, including entity attribute feature vectors of a mine historical knowledge graph, and modal features of multi-source observation spatio-temporal features (microseismic event feature sequence: time-ordered parameters such as "shock depth 520m, magnitude ML1.2", element concentration spatial distribution feature map: copper element 10m x 10m grid concentration map, and lithology-structure discrete feature sequence: "300-305m: basalt, no fracture" depth sequence); three structure-specialized cascade Transformer networks are constructed, the time series modal network has 1000 steps to adapt to microseismic sequences, the spatial modal network adds spatial attention, and the entity modal network optimizes the semantic embedding layer. The single-source layer uses an independent Transformer encoder, such as the entity layer that extracts the "F2 fault-eastern copper ore body" association; the shallow layer introduces cross-modal attention to map features to the same low-dimensional space and calculate microseismic and lithology feature weights to achieve alignment; the deep layer uses 8 cross-attention heads to extract the "copper enrichment area-lithology fragmentation-microseismic frequent occurrence" association from a causal perspective and reorganize high-dimensional semantic features; finally, the features are organized according to the mine area national 2000 three-dimensional coordinate system (longitude, latitude, and depth) to generate a deep-level fusion feature field.

[0054] Through the way provided by the embodiment, the carefully designed cascade Transformer network structure simulates the scientific cognitive process of geologists who first analyze each single data, then establish preliminary associations, and finally perform deep synthesis, making the fusion of multi-source data no longer a black box operation, but an intelligent processing flow with clear logic and traceable steps, significantly improving the credibility and interpretability of the fusion results.

[0055] In some embodiments, different inductive biases are deployed for different modal data directly related to mineralization potential, and a preliminary abstract feature screening is performed for mineralization: a graph structure enhanced Transformer encoder is used to process entity attribute feature vectors to strengthen the encoding of topological relationships between historical ore bodies, ore-controlling structures, and ore-forming favorable lithology entities, obtaining historical geological entity features; a time-aware Transformer encoder is used to process microseismic event feature sequences to extract their spatio-temporal clustering and energy release trend features, obtaining microseismic activity features; a space-aware Transformer encoder is used to process element concentration spatial distribution feature maps to identify the spatial morphology and concentration gradient features of element combination anomalies, obtaining geochemical anomaly features.

[0056] The preliminary abstract feature can be a feature that focuses on core information related to mineralization, which is obtained by a Transformer encoder with a mineralization-oriented inductive bias in a single-source feature extraction layer for different modal data directly related to mineralization potential. The historical ore body can be a mineral resource entity discovered and recorded in the historical exploration and mining process of the target mining area, with a clear spatial position, scale, ore grade, and mineral type, which is a core historical basis for reflecting the mineralization potential of the mining area. The ore-controlling structure can be a geological structure entity that controls the formation, distribution, shape, and scale of the ore body. The ore-forming favorable lithology entity can be a rock type entity with physical and chemical properties conducive to the formation and enrichment of mineral resources, such as sandstone, granite, basalt with specific components, etc. The historical geological entity feature can be an abstract feature of the historical geological entity (historical ore body, ore-controlling structure, ore-forming favorable lithology) focusing on mineralization obtained by processing entity attribute feature vectors with a graph structure enhanced Transformer encoder. The spatio-temporal clustering can be the concentrated distribution characteristics of microseismic events in the time and space dimensions. The energy release trend feature can be the variation law of the energy released by microseismic events over time, common trends include gradual energy enhancement, periodic energy fluctuation, and stable energy maintenance at a high level, which can indirectly reflect the activity intensity and stability of the ore-controlling structure. The microseismic activity feature can be an abstract feature obtained by processing microseismic event feature sequences with a time-aware Transformer encoder, which integrates the spatio-temporal clustering and energy release trend features of microseismic events. The geochemical anomaly feature can be an abstract feature obtained by processing element concentration spatial distribution feature maps with a space-aware Transformer encoder, which focuses on the combination anomaly information of ore-related elements (such as copper, lead, zinc, and sulfur).

[0057] Specifically, the traditional single-source feature extraction uses a general Transformer encoder without ore-forming oriented inductive bias, which cannot capture the core rules related to ore formation: when processing entity attribute feature vectors, the topological relationship of historical ore bodies, ore-controlling structures, etc. cannot be encoded; when processing microseismic event feature sequences, the spatiotemporal clustering and energy release trend are missed; when processing element concentration maps, it is difficult to identify the spatial form of ore-forming element combination anomalies, and a large amount of redundant information (such as ordinary rock layer attributes, random microseismic noise) will be extracted, interfering with subsequent fusion. To solve the above problems, this step first obtains the ore-forming related modal data: entity attribute feature vectors of the mine historical knowledge graph (such as “F3 ore-controlling fault” with a strike of N45° and a dip angle of 60°, “No. 5 ore body” with a length of 800m, and “Jurassic ore-forming sandstone” with a thickness of 50m), microseismic event feature sequences (such as “seismic depth 450m, magnitude ML1.3, coordinates E118°42′N30°26′” and “seismic depth 460m, magnitude ML1.5, coordinates E118°43′N30°27′” sorted by time), and element concentration spatial distribution feature maps (copper, lead, zinc 10m x 10m grid maps, such as copper 350mg / kg and lead 280mg / kg in a certain area). Then use three types of encoders for processing: a graph structure enhanced Transformer encoder converts entities into graphs (F3 fault and No. 5 ore body are set as “control relationship” edges), encodes topological relationships to get historical geological entity features; a time series perception encoder uses a 12-hour sliding window to extract microseismic spatiotemporal clustering (such as the above two events concentrated in the E118°42′-43′ area within 1 hour) and energy trend (magnitude from 1.3 to 1.5), to get microseismic activity features; a spatial perception encoder identifies element anomaly belt shape (such as the E118°40′-45′ area) and concentration gradient (center copper 800mg / kg decreases to 400mg / kg outside 50m) according to the ore-forming criterion (copper > 500mg / kg and lead > 300mg / kg), to get geological and chemical anomaly features; finally, the three types of features are standardized into 256-dimensional vectors for subsequent interaction.

[0058] By the way provided by the embodiment, by deploying specialized Transformer encoders with corresponding inductive bias for different data structures, the model architecture itself is more consistent with the inherent characteristics of various geological data, thereby achieving more in-depth and more essential information mining in the basic feature extraction stage, avoiding information loss or distortion caused by one-size-fits-all processing.

[0059] In some embodiments, a spatial position key lightweight cross-modal attention module is introduced to perform preliminary cross-validation and positioning of ore prospecting clues: the preliminary abstract features of each modality are projected into a unified gridded space based on their three-dimensional spatial coordinates, and the mutual attention weights between different modal features are analyzed through a lightweight cross-attention mechanism; based on the mutual attention weights, the historical geological entity features, microseismic activity features and geochemical anomaly features are weighted and fused to generate spatially coupled anomaly-structure joint features to delineate the ore-forming favorable area.

[0060] The lightweight cross-modal attention module can be a cross-modal attention calculation unit that uses three-dimensional spatial coordinates as the core association basis and has simplified model parameters and computational load. The cross-attention mechanism can be derived from data science, which refers to using the features of one modality (such as microseismic activity) to verify or explain the reliability and geological significance of the features of another modality (such as geochemical anomalies). The mutual attention weight can be the value of the importance of different modal features at the same grid node calculated by the lightweight cross-modal attention module (range 0-1), and a higher weight indicates that the modality feature has a stronger ore-forming indication significance for the current grid node. The anomaly-structure joint feature can be a coupled feature that contains information on geochemical anomalies (material traces), historical geological structures (spatial basis), and microseismic activity (activation state) after weighted fusion, which can preliminarily represent the ore-forming favorable conditions of "structure + anomaly + activation". The ore-forming favorable area can be a spatial region within the mining area that simultaneously possesses multiple favorable conditions such as ore-forming spatial basis (structure), material trace (element anomaly), and activation signal (microseismic), which is the key target area for mineral resource exploration.

[0061] Specifically, in the interaction process, the traditional multi-modal fusion method often uses a standard and global attention mechanism or simply concatenates features at the feature level, which has significant defects in geological prospecting applications. The former has a very high computational cost and is difficult to handle massive three-dimensional spatial data. The latter completely ignores the core geological constraint of the correspondence of the location of different modal features in the real three-dimensional space, and is prone to false associations, such as forcibly associating a deep microseismic event with a shallow and spatially unrelated geochemical anomaly, resulting in low clue reliability and fuzzy positioning of the prospecting target area. To address the above problems, the mine area is divided into a 10m x 10m x 5m grid according to the State 2000 coordinate system (such as node G001: E118°40'-40'10", N30°25'-25'10", depth 200-205m); the historical geological entity features (such as F3 ore-controlling fault) are projected onto the grid through which they pass, the microseismic activity features (2 microseismic events in the G001 area within 1 hour, magnitude ML1.1 rising to 1.3), and the geochemical anomaly features (copper 620 mg / kg, lead 350 mg / kg in the G001 area, meeting the ore-forming criteria) are projected onto the corresponding nodes; the lightweight cross-modal attention module is started, and the G001 is assigned a weight according to the ore-forming logic (structure 0.4, microseismic 0.3, element 0.3), and the weight of the node with only a single modal anomaly (such as G050 with only copper 550 mg / kg) is reduced to 0.1; the weighted fusion is performed according to the weight (G001 fusion feature = structure x 0.4 + microseismic x 0.3 + element x 0.3), and the missing information of the single modal is supplemented; a fusion feature threshold of 0.6 is set, and nodes greater than 0.6 are selected to generate abnormal-structure joint features.

[0062] By the way provided by the embodiment, the features of different modalities are forced to align to a unified spatial grid, and the interaction is performed based on local attention, effectively realizing mutual verification and supplement of multi-source information in the real geological space, filtering out spatially inconsistent noise signals, and significantly improving the accuracy of identifying ore-forming favorable positions.

[0063] In some embodiments, the abnormal-structure joint feature coupled in space is taken as input, and a mineralization pattern feature vector extracted from a mine historical knowledge graph is introduced as a global query; through a multi-head cross-attention mechanism, the depth semantic matching degree of the mineralization pattern feature vector and each spatial position joint is analyzed; the depth semantic matching degree is taken as the fusion feature vector of each spatial position in the deep fusion feature field, and the fusion feature vector directly represents the similarity between this spatial position and the historical mineralization pattern, i.e., the potential probability of remaining ore occurrence.

[0064] The ore-forming pattern feature vector can be extracted from the mine historical knowledge graph, representing the vector form data of "key condition combination of ore-forming", containing the historical verified core elements of ore-forming (such as the combination of "ore-controlling fault + ore-forming element anomaly + tectonic activation signal"), which is the condensed expression of ore-forming regularity. The deep semantic matching degree can be calculated by the multi-head cross attention mechanism, representing the similarity (numerical range 0-1) between the abnormal-structure joint features of a certain spatial position and the ore-forming pattern feature vector in the high-dimensional semantic layer. The higher the matching degree, the closer the ore-forming conditions of the position to the historical ore-forming pattern, and the higher the probability of remaining ore occurrence. The fusion feature vector can be a vector data integrating the spatial position ore-forming related information with the deep semantic matching degree as the core, directly representing the similarity between the spatial position and the historical ore-forming pattern, and essentially being the potential probability quantization result of the remaining ore occurrence. The potential probability of remaining ore occurrence can refer to the possibility of the existence of unexplored remaining mineral resources in a certain spatial position, which is directly represented by the fusion feature vector (deep semantic matching degree) and is the core quantitative index of resource potential evaluation.

[0065] Specifically, although the shallow layer interaction completes the local spatial cross-validation of multi-source clues, it still cannot simulate the core cognitive process of geological experts examining the metallogenic regularity from a global and systematic perspective. Traditional technologies either rely on expert experience for subjective inference or use conventional neural networks for limited nonlinear fusion, both of which are difficult to effectively capture the "long-range dependence" relationship in the metallogenic system, that is, the complex semantic association between different modal anomalies that may be far apart in space (such as hundreds of meters) but closely related in origin (such as deep rock mass and remote alteration zone). These global and high-level association patterns are the key to accurately delineating concealed ore bodies and understanding the entire metallogenic system. Without this deep fusion, the prediction will remain at the mechanical superposition level of local evidence and cannot generate a truly high-credibility quantitative prediction map that reflects the overall structure of the metallogenic system. To address the above problems, the step first obtains a spatially coupled anomaly-structure joint feature (such as grid node G001: F3 ore-controlling fault strikes N45°, copper 620 mg / kg, lead 350 mg / kg, 2 ML1.2-ML1.3 microseisms within 1 hour), and then extracts a metallogenic pattern feature vector V (such as "fault N30°-N60°+ copper > 500 mg / kg+ microseism ≥ ML1.2", quantified as 256 dimensions) from the mine historical knowledge graph; configure 6 cross-attention modules (each head focuses on fault strike, element concentration, microseismic intensity, etc.), take V as the global query, and G001 feature as the key / value, calculate the matching degree of each head (fault 0.9, element 0.85, microseismic 0.9, etc.), and the average value 0.88 is the deep semantic matching degree; integrate the matching degree and the 3D coordinates of G001 (E118°40', N30°25', 203m) to generate a fusion feature vector, embed the deep-level fusion feature field according to the coordinates; verify the node G500 without structure, and its matching degree 0.12 determines it as a low ore probability area, ensuring the reliability of the feature.

[0066] In the manner provided by the embodiment, the final output metallogenic potential probability map visually and quantitatively displays the spatial distribution of metallogenic possibility, with high resolution and strong reliability, and can be directly used as a core scientific basis for selecting exploration targets and deploying drilling projects, greatly improving the accuracy and efficiency of mineral exploration.

[0067] In some embodiments, based on the deep-level fusion feature field, double-path feature representations for disaster risk prediction and resource potential evaluation are extracted in parallel; in the disaster risk quantification path, the deep-level fusion feature field is constrainedly fused with the geomechanics prior rules, and the coupling calculation of the concealed structure activation degree and the rock mass stability parameters is performed to generate a disaster risk probability value for each voxel; in the resource economic value quantification path, the fusion feature vector representing the ore-forming model matching degree is coupled with the ore grade, mining cost and market parameters for multi-objective optimization to generate a resource economic value index for each voxel; the disaster risk probability value and the resource economic value index are integrated at the voxel level in a unified three-dimensional geological space framework to generate a risk-resource integrated three-dimensional geological model containing both risk and value quantification attributes for each spatial location.

[0068] The disaster risk quantification path can be a processing flow focusing on "calculating the probability of geological disaster occurrence", and the core is to convert the geological stability features into quantifiable disaster risk indicators by combining geomechanics rules, covering the coupling calculation of the concealed structure activation degree and the rock mass stability parameters. A voxel can be the smallest volume unit in a three-dimensional space with a unified attribute value (such as density, probability), similar to a pixel in a two-dimensional image. The concealed structure activation degree can be a quantitative indicator (numerical range 0-1) representing whether the unexposed (concealed) geological structure (such as fault, fracture) in the mining area is in an active state, and the higher the activation degree, the greater the risk of triggering disasters such as landslides and collapses due to tectonic activity. The rock mass stability parameter can be a quantitative indicator representing the rock mass's resistance to deformation and destruction, including rock mass integrity coefficient, compressive strength, water content, etc., which directly affects the possibility of disaster occurrence. The disaster risk probability value can be the possibility of a certain voxel node occurring geological disaster (numerical range 0-1, such as 0.8 representing high risk and 0.2 representing low risk) obtained after coupling calculation of the concealed structure activation degree and the rock mass stability parameter, which is the final output of the disaster risk quantification path. The resource economic value quantification path can be a processing flow focusing on "evaluating the economic benefits of mineral resources", and the core is to convert the ore-forming value features into quantifiable economic value indicators by combining the ore-forming model matching degree, mining cost and market conditions. The resource economic value index can be the economic value quantification result of the mineral resources in the region corresponding to a certain voxel node (such as 1200 yuan / cubic meter), and the higher the value, the greater the economic benefit potential of the resources in the region, which is the final output of the resource economic value quantification path. The voxel-level integration can be to bind the disaster risk probability value and the resource economic value index corresponding to each voxel in a unified three-dimensional framework, so that each voxel has both "risk-value" attributes, realizing the integrated correlation at the microscopic level.

[0069] Specifically, the traditional geological model separates disaster risk assessment from resource value assessment, which easily leads to decision-making contradictions such as "ignoring high risk in high value area" or "over management in low value area"; disaster quantification only relies on single data such as microseismic frequency, without the constraint of geomechanical prior rules, and the result deviates from the stress law of rock mass; resource value assessment only considers ore grade, ignores the matching degree of ore-forming model (residual ore probability), and is easily overestimated; and only regional integration is available, which is insufficient in precision. To solve the above problems, this step extracts geological stability features (such as voxel G001: F3 fault, 2 ML1.2-ML1.3 microseisms per hour, basalt) and ore-forming value features (G001 ore-forming matching degree 0.88, copper grade 1.2%) from the deep level of fusion feature field (voxel resolution 10m x 10m x 5m) in parallel; disaster path fusion geomechanical rules (such as activation degree = microseismic energy x rock mass fragmentation coefficient), calculate G001 activation degree 0.3, rock mass stability 0.75, and get disaster risk probability 0.28 according to the formula; resource path according to weight (matching degree 0.4, grade 0.3, etc.) G001 score 0.892, combined with copper price 60,000 yuan / ton, mining cost 80 yuan / ton, and tax 13%, value index 1260 yuan / m3; bind voxel risk and value (G001: 0.28+1260), render by color scale (risk red = 0.7-1.0, value gold = 1000 yuan / m3+), generate a clickable query risk-resource integrated 3D geological model.

[0070] In the manner provided by this embodiment, the continuous model probability output is converted into a target area map with clear boundaries and explicit priority classification, so that geological exploration personnel can directly apply it to engineering deployment without complex interpretation, realizing seamless connection between artificial intelligence prediction and field practice, and significantly improving the efficiency of formulating exploration schemes.

[0071] Figure 3 The structure diagram of the cascade Transformer fusion system based on multi-source heterogeneous geological data provided by an embodiment of the present application is shown in Figure 3 The cascade Transformer fusion system based on multi-source heterogeneous geological data of the embodiment includes a geological feature analysis module 301, a feature field generation module 302, a three-dimensional module construction module 303, and an optimization scheme generation module 304. The geological feature analysis module 301 is configured to obtain historical paper geological data sets and real-time multi-source geological monitoring data streams, perform structural analysis and space-time coding on the historical paper geological data sets, and perform feature extraction and serialization on the real-time multi-source geological monitoring data streams, to generate a multi-source heterogeneous geological feature set; the feature field generation module 302 is configured to generate a deep-level fusion feature field by performing multi-level feature extraction and cross-modal attention fusion from coarse to fine through a cascade Transformer network based on the multi-source heterogeneous geological feature set; the three-dimensional module construction module 303 is configured to generate a risk-resource integrated three-dimensional geological model of quantified disaster risk probability and resource economic value by performing multi-objective collaborative optimization calculation based on the deep-level fusion feature field and incorporating geological priori knowledge rules; and the optimization scheme generation module 304 is configured to generate a set of collaborative optimization engineering schemes by performing optimization of hidden disaster targeted governance and resource re-mining through a decision mapping engine based on the risk-resource integrated three-dimensional geological model.

[0072] Optionally, in the generation of the multi-source heterogeneous geological feature set, the geological feature analysis module 301 is specifically configured to: perform structural analysis and coding based on optical character recognition, visual models and space-time correlation on the historical paper geological data sets, to generate a structured mine historical knowledge graph; perform multi-modal feature extraction and space-time alignment processing on the real-time multi-source geological monitoring data streams, to generate a unified benchmark multi-source observation space-time feature tensor; and generate the multi-source heterogeneous geological feature set based on an entity attribute feature vector extracted from the mine historical knowledge graph and the multi-source observation space-time feature tensor.

[0073] Optionally, in the generation of the structured mine historical knowledge graph, the geological feature analysis module 301 is specifically configured to: the historical paper geological data sets include old mine geological exploration report texts, old mine hand-drawn geological section maps, old mine drilling log tables and historical mining record tables; extract standardized geological entities and geological semantic relationships between the geological entities from the old mine geological exploration report texts through cascade processing of an optical character recognition model and geological field professional terms; identify line symbols in the old mine hand-drawn geological section maps through a visual transformer, and analyze and reconstruct the line symbols into vectorized graphic objects with spatial coordinates and geological attributes; perform space-time correlation alignment of text structured data in the old mine drilling log tables and the historical mining record tables with entities extracted from the old mine geological exploration report texts and the old mine hand-drawn geological section maps; and construct the mine historical knowledge graph based on the geological semantic relationships, the vectorized graphic objects and the aligned text structured data.

[0074] Optionally, the geological feature analysis module 301, when generating the multi-source observation spatio-temporal feature tensor based on the generated unified reference, is specifically configured to: the real-time multi-source geological monitoring data stream includes microseismic monitoring waveform time series data, geological chemical element sampling point data, and engineering drilling core logging data; extract the source parameters from the microseismic monitoring waveform time series data based on an event detection algorithm to form a microseismic event feature sequence in units of events; process the geological chemical element sampling point data based on a spatial interpolation and outlier analysis algorithm to generate an element concentration spatial distribution feature map; analyze the engineering drilling core logging data based on a natural language processing and lithology classification model to generate a lithology-structure discrete feature sequence with depth coordinates; and convert the microseismic event feature sequence, the element concentration spatial distribution feature map, and the lithology-structure discrete feature sequence to the same three-dimensional geographic coordinate system and time reference to generate the multi-source observation spatio-temporal feature tensor.

[0075] Optionally, the feature field generation module 302, when generating the deep-level fusion feature field based on the generated multi-source heterogeneous geological feature set, is specifically configured to: input the multi-source heterogeneous geological feature set into a plurality of independent and structure-specialized cascaded Transformer networks, the cascaded Transformer networks including single-source feature extraction layers, shallow cross-modal interaction layers, and deep cross-modal fusion layers connected in sequence; in the single-source feature extraction layers, independently abstract feature analyze the entity attribute feature vector from the mine historical knowledge graph and the different modal feature sequences from the multi-source observation spatio-temporal feature tensor through independent Transformer encoders; in the shallow cross-modal interaction layers, introduce a cross-modal attention mechanism to perform low-dimensional interaction and alignment on the preliminarily abstracted modal features; and in the deep cross-modal fusion layers, perform high-dimensional semantic feature reorganization and deep fusion through a multi-head cross-attention mechanism to output the deep-level fusion feature field representing the geological spatio-temporal structure and attributes.

[0076] Optionally, the feature field generation module 302, based on the single-source feature extraction layer, is specifically configured to: for different modal data directly related to ore-forming potential, deploy a Transformer encoder with different inductive bias to conduct preliminary abstract feature screening for ore formation: use a graph structure enhanced Transformer encoder to process the entity attribute feature vector to strengthen the encoding of the topological relationship between historical ore bodies, ore-controlling structures, and ore-forming favorable lithology entities, obtaining historical geological entity features; use a time-aware Transformer encoder to process the microseismic event feature sequence to extract its spatiotemporal clustering and energy release trend features, obtaining microseismic activity features; use a space-aware Transformer encoder to process the element concentration spatial distribution feature map to identify the spatial form and concentration gradient features of element combination anomalies, obtaining geochemical anomaly features.

[0077] Optionally, the feature field generation module 302, based on the shallow cross-modal interaction layer, is specifically configured to: introduce a lightweight cross-modal attention module based on spatial location to perform preliminary cross-validation and positioning focus of ore prospecting clues: project the preliminary abstract features of each modality into a unified gridded space based on their three-dimensional spatial coordinates, and analyze the mutual attention weights between different modal features through a lightweight cross-attention mechanism; based on the mutual attention weights, perform weighted fusion and information complementation on the historical geological entity features, the microseismic activity features, and the geochemical anomaly features to generate spatially preliminarily coupled anomaly-structure joint features to delineate ore-forming favorable areas.

[0078] Optionally, the feature field generation module 302, based on the deep cross-modal fusion layer, is specifically configured to: take the spatially coupled anomaly-structure joint features as input and introduce a mineralization pattern feature vector refined from the mine historical knowledge graph as a global query; analyze the deep semantic matching degree of the mineralization pattern feature vector and each spatial location joint through a multi-head cross-attention mechanism; take the deep semantic matching degree as the fusion feature vector of each spatial location in the deep fusion feature field, which directly represents the similarity between this spatial location and historical mineralization patterns, i.e., the potential probability of residual ore occurrence.

[0079] Optionally, the three-dimensional module construction module 303 is specifically configured to: based on the deep-level fusion feature field, parallel extraction of double-path feature representation for disaster risk prediction and resource potential assessment; in the disaster risk quantification path, the deep-level fusion feature field is constrainedly fused with the geomechanics prior rule, the disaster risk probability value is generated voxel by voxel through the coupling calculation of the concealed structure activation degree and the rock mass stability parameter; in the resource economic value quantification path, the fusion feature vector representing the ore-forming pattern matching degree is coupled with the ore grade, the mining cost and the market parameter for multi-objective optimization, and the resource economic value index is generated voxel by voxel; the disaster risk probability value and the resource economic value index are integrated at the voxel level under the unified three-dimensional geological space framework, and a risk-resource integrated three-dimensional geological model containing the risk and value quantification attributes of each spatial position is generated.

[0080] The system of the embodiment can be used to execute the method of any of the above embodiments, and has similar implementation principles and technical effects, which will not be described here again.

Claims

1. A cascading Transformer fusion method based on multi-source heterogeneous geological data, characterized in that, The method comprises the following steps: acquiring historical paper geological data sets and real-time multi-source geological monitoring data streams, structurally analyzing and spatio-temporally coding the historical paper geological data sets, and extracting features and sequencing the real-time multi-source geological monitoring data streams to generate a multi-source heterogeneous geological feature set; based on the multi-source heterogeneous geological feature set, performing coarse-to-fine multi-level feature extraction and cross-modal attention fusion through a cascade Transformer network to generate a deep-level fusion feature field; based on the deep-level fusion feature field, integrating geological priori knowledge rules to perform multi-objective collaborative optimization calculation to generate a risk-resource integrated three-dimensional geological model of quantitative disaster risk probability and resource economic value; based on the risk-resource integrated three-dimensional geological model, performing hidden disaster targeted management and resource re-mining optimization through a decision mapping engine to generate a set of collaborative optimization engineering schemes.

2. The method of claim 1, wherein, The method of generating a multi-source heterogeneous geological feature set comprises: structurally analyzing and coding the historical paper geological data sets based on optical character recognition, visual models and spatio-temporal correlation to generate a structured mine historical knowledge graph; performing multi-modal feature extraction and spatio-temporal alignment processing on the real-time multi-source geological monitoring data streams to generate a unified benchmark multi-source observation spatio-temporal feature tensor; based on the entity attribute feature vector extracted from the mine historical knowledge graph and the multi-source observation spatio-temporal feature tensor, generating the multi-source heterogeneous geological feature set.

3. The method of claim 2, wherein, The method of generating a structured mine historical knowledge graph comprises: the historical paper geological data set comprises old mine geological exploration report texts, old mine hand-drawn geological section maps, old mine drilling log tables and historical mining record tables; extracting standardized geological entities and geological semantic relationships between the geological entities from the old mine geological exploration report texts through a cascade processing of an optical character recognition model and geological field professional terms; identifying line symbols in the old mine hand-drawn geological section maps through a visual transformer, and analyzing and reconstructing them into vectorized graphic objects with spatial coordinates and geological attributes; spatio-temporally correlating and aligning the text structured data in the old mine drilling log tables and the historical mining record tables with the entities extracted from the old mine geological exploration report texts and the old mine hand-drawn geological section maps; based on the geological semantic relationships, the vectorized graphic objects and the aligned text structured data, constructing the mine historical knowledge graph.

4. The method of claim 2, wherein, The method of generating a unified benchmark multi-source observation spatio-temporal feature tensor comprises: the real-time multi-source geological monitoring data stream comprises microseismic monitoring waveform time series data, geological chemical element sampling point data and engineering drilling core logging data; extracting source parameters from the microseismic monitoring waveform time series data based on an event detection algorithm to form a microseismic event feature sequence in units of events; processing the geological chemical element sampling point data based on a spatial interpolation and outlier analysis algorithm to generate an element concentration spatial distribution feature map; analyzing the engineering drilling core logging data based on a natural language processing and lithology classification model to generate a lithology-structure discrete feature sequence with depth coordinates; The microseismic event feature sequence, the element concentration spatial distribution feature map, and the lithology-structure discrete feature sequence are uniformly converted to the same three-dimensional geographic coordinate system and time reference to generate the multi-source observation spatiotemporal feature tensor.

5. The method of claim 4, wherein, The deep-level fusion feature field is generated by: inputting the multi-source heterogeneous geological feature set into a plurality of independent and structure-specialized cascaded Transformer networks, the cascaded Transformer networks comprising single-source feature extraction layers, shallow cross-modal interaction layers, and deep cross-modal fusion layers connected in sequence; in the single-source feature extraction layer, independently performing preliminary abstract feature analysis on the entity attribute feature vector from the mine historical knowledge graph and different modal feature sequences from the multi-source observation spatiotemporal feature tensor by using independent Transformer encoders; in the shallow cross-modal interaction layer, introducing a cross-modal attention mechanism to perform low-dimensional interaction and alignment on the preliminarily abstracted modal features; in the deep cross-modal fusion layer, performing high-dimensional semantic feature reorganization and deep fusion by using a multi-head cross-attention mechanism, and outputting the deep-level fusion feature field representing the geological spatiotemporal structure and attributes.

6. The method of claim 5, wherein, The single-source feature extraction layer includes: deploying Transformer encoders with different induction biases for different modal data directly related to mineralization potential to perform the preliminary abstract feature screening for mineralization: using a graph structure enhanced Transformer encoder to process the entity attribute feature vector to strengthen the encoding of the topological relationship among historical ore bodies, ore-controlling structures, and ore-forming favorable lithology entities, and obtain historical geological entity features; using a time sequence aware Transformer encoder to process the microseismic event feature sequence to extract its spatiotemporal aggregation and energy release trend features, and obtain microseismic activity features; using a space aware Transformer encoder to process the element concentration spatial distribution feature map to identify the spatial form and concentration gradient features of element combination anomalies, and obtain geochemical anomaly features.

7. The method of claim 6, wherein, The shallow cross-modal interaction layer includes: introducing a lightweight cross-modal attention module based on spatial position to perform initial cross-validation and positioning focusing of ore-prospecting clues: projecting the preliminary abstract features of each modal into a unified gridded space based on their three-dimensional spatial coordinates, and analyzing the mutual attention weights between different modal features by using a lightweight cross-attention mechanism; based on the mutual attention weights, performing weighted fusion and information complementation on the historical geological entity features, the microseismic activity features, and the geochemical anomaly features to generate spatially preliminarily coupled anomaly-structure joint features to delineate ore-forming favorable areas.

8. The method of claim 7, wherein, The deep cross-modal fusion layer includes: taking the spatially coupled anomaly-structure joint features as input and introducing a mineralization pattern feature vector refined from the mine historical knowledge graph as a global query; analyzing the deep semantic matching degree between the mineralization pattern feature vector and each spatial position joint by using a multi-head cross-attention mechanism; and The deep semantic matching degree is taken as a fusion feature vector of each spatial position in the deep-level fusion feature field, and the fusion feature vector directly represents the similarity between the spatial position and the historical ore-forming pattern, i.e., the potential probability of the remaining ore occurrence.

9. The method of claim 8, wherein, The generated risk-resource integrated three-dimensional geological model of quantitative disaster risk probability and resource economic value comprises: Based on the deep-level fusion feature field, a dual-path feature representation for disaster risk prediction and resource potential evaluation is extracted in parallel; In the disaster risk quantification path, the deep-level fusion feature field is constrainedly fused with the geomechanics prior rules, and the disaster risk probability value is generated voxel by voxel through the coupling calculation of the concealed structure activation degree and the rock mass stability parameters; In the resource economic value quantification path, the fusion feature vector representing the ore-forming pattern matching degree is coupled with the ore grade, mining cost and market parameters for multi-objective optimization, and the resource economic value index is generated voxel by voxel; The disaster risk probability value and the resource economic value index are integrated at the voxel level under a unified three-dimensional geological space framework to generate a risk-resource integrated three-dimensional geological model of each spatial position containing both risk and value quantitative attributes.

10. A cascaded Transformer fusion system based on multi-source heterogeneous geological data, characterized in that, The application is applied to the method of any one of claims 1-9, comprising: a geological feature analysis module for obtaining a historical paper geological data set and a real-time multi-source geological monitoring data stream, performing structured analysis and spatio-temporal coding on the historical paper geological data set, and performing feature extraction and serialization on the real-time multi-source geological monitoring data stream to generate a multi-source heterogeneous geological feature set; a feature field generation module for generating a deep-level fusion feature field by performing multi-level feature extraction and cross-modal attention fusion from coarse to fine through a cascade Transformer network based on the multi-source heterogeneous geological feature set; a three-dimensional module construction module for generating a risk-resource integrated three-dimensional geological model of quantitative disaster risk probability and resource economic value by performing multi-objective collaborative optimization calculation based on the deep-level fusion feature field and incorporating geology prior knowledge rules; an optimization scheme generation module for generating a collaborative optimization engineering scheme set by performing optimization of concealed disaster targeted governance and resource re-mining through a decision mapping engine based on the risk-resource integrated three-dimensional geological model.

Images