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

Abstract

The 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. The method comprises the following steps: acquiring a historical paper geological data set and real-time multi-source geological monitoring data flow, based on which a multi-source heterogeneous geological feature set is generated; based on this, coarse-to-fine multi-level feature extraction and cross-modal attention fusion are carried out through a cascading Transformer network to generate a deep-level fusion feature field; based on this, multi-objective collaborative optimization calculation is carried out by integrating geological prior knowledge rules to generate a risk-resource integrated three-dimensional geological model of quantitative disaster risk probability and resource economic value; based on this, the optimization of hidden disaster targeted management and resource re-mining is carried out through a decision mapping engine to generate a set of collaborative optimization engineering schemes. In the data cascading fusion process, the intelligent level and comprehensive benefits of geological engineering activities are significantly 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 residual ore and disaster risk early 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 mines, 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, which 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:

[0006] Obtaining a historical paper geological data set and a real-time multi-source geological monitoring data stream, structurally analyzing and spatiotemporally encoding the historical paper geological data set, and extracting features and sequencing the real-time multi-source geological monitoring data stream to generate a multi-source heterogeneous geological feature set;

[0007] Based on the multi-source heterogeneous geological feature set, 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;

[0008] 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;

[0009] 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.

[0010] By the technical scheme, based on automatic and intelligent scheme generation, the efficiency and scientificity of engineering planning are improved, deep cooperation of disaster control and resource development in space, time sequence and benefit is realized, and an intelligent decision tool is provided for mine safety production, safe use of urban underground space, and major engineering site selection.

[0011] Optionally, the generating the multi-source heterogeneous geological feature set comprises:

[0012] Performing structured analysis and coding on the historical paper geological data set based on optical character recognition, visual model and space-time correlation to generate a structured mine history knowledge graph;

[0013] Performing multi-modal feature extraction and space-time alignment processing on the real-time multi-source geological monitoring data stream to generate a unified reference multi-source observation space-time feature tensor;

[0014] Based on the entity attribute feature vector extracted from the mine history knowledge graph and the multi-source observation space-time feature tensor, the multi-source heterogeneous geological feature set is generated.

[0015] Optionally, the generating the structured mine history knowledge graph comprises:

[0016] 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;

[0017] Through cascade processing of an optical character recognition model and geological field professional terms, standardized geological entities and geological semantic relationships between the geological entities are extracted from the old mine geological exploration report texts;

[0018] Through a visual transformer, line symbol in the old mine hand-drawn geological section map is recognized, and is analyzed and reconstructed into a vectorized graphic object with spatial coordinates and geological attributes;

[0019] The text structured data in the old mine drilling log table and the historical mining record table is time-space correlated and aligned with the entities extracted from the old mine geological exploration report texts and the old mine hand-drawn geological section map;

[0020] Based on the geological semantic relationships, the vectorized graphic object and the aligned text structured data, the mine history knowledge graph is constructed.

[0021] Optionally, the generating the unified reference multi-source observation space-time feature tensor comprises:

[0022] The real-time multi-source geological monitoring data stream comprises microseismic monitoring waveform time sequence data, geological chemical element sampling point data and engineering drilling core logging data.

[0023] 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;

[0024] 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;

[0025] 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;

[0026] unifying 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 spatiotemporal feature tensor.

[0027] Optionally, the generating a deep-level fusion feature field comprises:

[0028] 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;

[0029] 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 through Transformer encoders;

[0030] 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;

[0031] 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 spatiotemporal structure and attributes.

[0032] Optionally, the single-source feature extraction layer comprises:

[0033] deploying Transformer encoders with different inductive biases for different modal data directly related to ore-forming potential to perform the preliminary abstract feature screening for ore formation:

[0034] The entity attribute feature vectors are processed using a graph-enhanced Transformer encoder to strengthen the encoding of topological relationships between historical ore bodies, ore-controlling structures, and mineralization-favorable lithological entities, thereby obtaining historical geological entity features.

[0035] The microseismic event feature sequence is processed using a time-aware Transformer encoder to extract its spatiotemporal clustering and energy release trend features, thereby obtaining microseismic activity features.

[0036] The spatially aware Transformer encoder is used to process the spatial distribution feature map of element concentrations to identify the spatial morphology and concentration gradient features of element combination anomalies, thereby obtaining geochemical anomaly features.

[0037] Optionally, the shallow cross-modal interaction layer includes:

[0038] A lightweight cross-modal attention module with spatial location as the key is introduced to perform initial cross-validation and localization focusing of mineral exploration clues:

[0039] The preliminary abstract features of each modality are projected onto a unified gridded space based on their three-dimensional spatial coordinates, and the mutual attention weights between features of different modalities are analyzed through a lightweight cross-attention mechanism.

[0040] Based on the aforementioned mutual attention weights, the historical geological entity characteristics, the microseismic activity characteristics, and the geochemical anomaly characteristics are weighted, fused, and complemented to generate spatially preliminarily coupled anomaly-tectonic joint features, thereby delineating favorable mineralization areas.

[0041] Optionally, the deep cross-modal fusion layer includes:

[0042] Using the spatially coupled anomaly-structure joint features as input, the metallogenic pattern feature vector extracted from the mine history knowledge graph is introduced as a global query.

[0043] The deep semantic matching degree between the feature vector of the mineralization mode and each of the spatial locations is analyzed by using a multi-head cross-attention mechanism.

[0044] The deep semantic matching degree is used as the fusion feature vector of each spatial location in the deep fusion feature field. The fusion feature vector directly represents the similarity between this spatial location and the historical mineralization pattern, that is, the potential probability of residual mineralization.

[0045] Optionally, the generation of the risk-resource integrated three-dimensional geological model that quantifies the probability of disaster risk and the economic value of resources includes:

[0046] Based on the deep fusion feature field, dual-path feature representations for disaster risk prediction and resource potential assessment are extracted in parallel;

[0047] In the disaster risk quantification path, the deep-level fusion feature field is constrainedly fused with the geomechanical a priori rules, and disaster risk probability values ​​are generated voxel by voxel through the coupling calculation of the activation degree of concealed structures and rock mass stability parameters.

[0048] In the path of quantifying the economic value of resources, the fusion feature vector representing the matching degree of mineralization model is coupled with ore grade, mining cost and market parameters through multi-objective optimization, and the resource economic value index is generated on a voxel-by-voxel basis.

[0049] The disaster risk probability value and the resource economic value index are integrated at the voxel level within a unified three-dimensional geological spatial framework to generate a risk-resource integrated three-dimensional geological model for each spatial location that simultaneously includes risk and value quantification attributes.

[0050] Secondly, this application provides a cascaded Transformer fusion system based on multi-source heterogeneous geological data, the system comprising:

[0051] The geological feature analysis module is used to acquire historical paper-based geological data sets and real-time multi-source geological monitoring data streams, perform structured parsing and spatiotemporal encoding on the historical paper-based geological data sets, and extract and serialize features from the real-time multi-source geological monitoring data streams to generate multi-source heterogeneous geological feature sets.

[0052] The feature field generation module is used to generate a deep-level fused feature field by performing multi-level feature extraction from coarse to fine and cross-modal attention fusion through a cascaded Transformer network based on the multi-source heterogeneous geological feature set.

[0053] The three-dimensional module construction module is used to perform multi-objective collaborative optimization calculations based on the deep-level fusion feature field and incorporate geological prior knowledge rules to generate a risk-resource integrated three-dimensional geological model that quantifies the probability of disaster risk and the economic value of resources.

[0054] The optimization scheme generation module is used to optimize the targeted management of hidden disasters and the re-exploitation of resources based on the risk-resource integrated three-dimensional geological model through a decision mapping engine, and generate a set of collaborative optimization engineering schemes. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a schematic diagram illustrating an application scenario provided in one embodiment of this application;

[0057] Figure 2 A flowchart illustrating a cascaded Transformer fusion method based on multi-source heterogeneous geological data provided in an embodiment of this application;

[0058] Figure 3 This is a schematic diagram of the structure of a cascaded Transformer fusion system based on multi-source heterogeneous geological data provided in an embodiment of this application. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0060] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0061] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0062] Existing geological data analysis and modeling methods lack a mechanism for unified characterization, dynamic fusion, and collaborative interpretation of deep-seated spatiotemporal correlations and physical coupling relationships between data when faced with multi-source, heterogeneous, and spatiotemporally non-uniform geological data. This not only fails to effectively support the accurate delineation of the location of hidden resources, but also fails to quantitatively assess potential disaster risks.

[0063] Based on this, this application provides a cascaded Transformer fusion method and system based on multi-source heterogeneous geological data. First, historical paper data and real-time monitoring data streams are accessed in parallel. NLP and signal processing techniques are used to perform structured analysis, spatiotemporal encoding, and feature serialization to construct a multi-source heterogeneous geological feature set within a unified spatiotemporal framework. Then, through a cascaded Transformer network and multi-level cross-modal attention computation, data-driven deep feature fusion from local details to global patterns is achieved, generating a deep-level fused feature field. Next, formalized geological prior knowledge rules are used as constraints, and a multi-objective collaborative optimization algorithm drives the generation network to output an integrated three-dimensional geological model that simultaneously quantifies disaster risk and resource value. Finally, based on this model, a decision mapping engine uses intelligent optimization algorithms to automatically search for and generate a set of Pareto-optimal engineering solutions that coordinate targeted management of hidden disasters and resource re-exploitation, and outputs this solution to geological researchers. This method overcomes the long-standing problems of data silos and dormant information in the geological field, generating a quantitative three-dimensional model that integrates risk and resources. It provides a unified scientific decision-making basis for the precise early warning and prevention of hidden disasters and the safe and green development of mineral resources, significantly improving the intelligence level and comprehensive benefits of geological engineering activities.

[0064] Figure 1 This is a schematic diagram illustrating an application scenario provided by this application. In the process of data cascading and fusion, this application utilizes the method provided to offer a scientific decision-making foundation for the safe and green development of mineral resources, significantly improving the intelligence level and comprehensive benefits of geological engineering activities.

[0065] Specifically, the method of this application is applied to any server that communicates with a mine history archive and a geological monitoring sensor network. The server acquires historical paper-based geological data from the mine history archive and real-time multi-source geological monitoring data streams from the geological monitoring sensor network. First, the historical paper-based data and real-time monitoring data streams are accessed in parallel. NLP and signal processing techniques are used to perform structured analysis, spatiotemporal coding, and feature serialization to construct a multi-source heterogeneous geological feature set within a unified spatiotemporal framework. Then, through a cascaded Transformer network and multi-level cross-modal attention computation, data-driven deep feature fusion from local details to global patterns is achieved, generating a deep-level fused feature field. Furthermore, formalized geological prior knowledge rules are used as constraints, and a multi-objective collaborative optimization algorithm drives the generation network to output an integrated three-dimensional geological model that simultaneously quantifies disaster risk and resource value. Finally, based on this model, a decision mapping engine uses intelligent optimization algorithms to automatically search for and generate a set of Pareto-optimal engineering solutions that coordinate targeted management of hidden disasters and resource re-mining, and outputs this set to geological researchers.

[0066] For specific implementation details, please refer to the following examples.

[0067] Figure 2 This is a flowchart illustrating a cascaded Transformer fusion method based on multi-source heterogeneous geological data according to an embodiment of this application. The method of this embodiment can be applied to servers in the above scenario. Figure 2 As shown, the method includes:

[0068] S201. Obtain historical paper-based geological data sets and real-time multi-source geological monitoring data streams. Perform structured analysis and spatiotemporal coding on the historical paper-based geological data sets, and extract and serialize features from the real-time multi-source geological monitoring data streams to generate a multi-source heterogeneous geological feature set.

[0069] Historical paper-based geological data sets can be collections of various geological-related data accumulated over a long period of time in paper form, including old mine geological exploration reports, hand-drawn geological profiles of old mines, old mine drilling logs, and historical mining records. The data originates from the company's historical mine archives. Real-time multi-source geological monitoring data streams can be continuous transmission streams of geological environment and geological state-related data collected in real time from different monitoring dimensions by various geological monitoring equipment. These include microseismic monitoring waveform time-series data, geochemical element sampling point data, and engineering drilling core logging data. The data originates from various geological monitoring sensor networks deployed within the monitoring area. Multi-source heterogeneous geological feature sets can integrate historical geological data features that have undergone structured analysis and spatiotemporal coding, as well as real-time geological monitoring data features that have undergone feature extraction and serialization.

[0070] Specifically, in the fields of geological exploration, disaster prevention, and resource development, existing technologies have long faced prominent problems such as insufficient utilization of data sources and poor data integration, which seriously restricts the comprehensiveness and accuracy of geological analysis. On the one hand, historical paper-based geological data is a valuable asset accumulated through long-term geological research and engineering practice, containing irreplaceable basic geological information. However, due to its paper-based format, it suffers from inherent defects such as scattered storage, difficulty in retrieval, and inability to be directly processed by computers. This results in these historical data remaining dormant for a long time, making it difficult to effectively complement real-time monitoring data and causing a serious waste of geological data resources. For example, a mining company's decades-old paper-based exploration archives contain key information such as the early stratigraphic distribution and structural characteristics of the area. However, because they cannot be quickly analyzed and utilized, they were not fully referenced in subsequent resource extraction planning, leading to deviations between the extraction plan and actual geological conditions, increasing extraction risks and costs. On the other hand, with the rapid development of geological monitoring technology, various monitoring devices have emerged, forming multi-source heterogeneous real-time data streams. These data come from different sources, have different formats, and are diverse in dimensions. Traditional data processing methods often can only perform simple processing on data of a single type or source, lacking effective mining of the correlation features in multi-source data, resulting in the inability to fully release the value of the data. This step collects historical paper geological data and real-time multi-source monitoring data streams. The paper data is scanned, extracted by OCR, and structured and parsed before being encoded with spatiotemporal codes. The real-time data is preprocessed, and core features are extracted and serialized. Finally, the two types of features are integrated to generate a multi-source heterogeneous geological feature set. By standardizing, structuring, and integrating historical geological data with real-time monitoring data in a spatiotemporal manner, dormant paper archives and empirical descriptions are transformed into computable data, and high-frequency, multi-dimensional monitoring data is refined into geologically meaningful feature sequences. This provides high-quality, semantically related input for subsequent deep information fusion, fundamentally changing the situation of scattered, heterogeneous, and difficult-to-integrate geological data.

[0071] S202. Based on a multi-source heterogeneous geological feature set, a multi-level feature extraction process from coarse to fine and cross-modal attention fusion are performed through a cascaded Transformer network to generate a deep-level fused feature field.

[0072] A cascaded Transformer network is a deep learning network model composed of multiple Transformer subnetworks connected in a hierarchical manner. Each subnetwork undertakes a different level of feature extraction task, and the output of the previous subnetwork serves as the input of the next subnetwork. Multi-level feature extraction can unfold the feature extraction process from the whole to the part, from the surface to the core. First, a primary Transformer subnetwork extracts shallow, general features (coarse-grained features) from a set of multi-source heterogeneous geological features. Then, subsequent Transformer subnetworks at each level gradually mine detailed information, correlation information, and potential patterns (fine-grained features) from the features, achieving a progressively deeper feature extraction. Cross-modal attention fusion can be used to calculate the correlation weights between features of different modalities (such as historical structured data modalities, real-time sensor data modalities, image data modalities, etc.) in a set of multi-source heterogeneous geological features through an attention mechanism. The deep-level fusion feature field can be a high-dimensional, integrated feature space that comprehensively and deeply reflects the geological history, real-time dynamic changes, and inherent correlation laws of various geological factors in the monitoring area after multi-level extraction and cross-modal fusion through cascaded Transformer networks. It includes features at all levels from coarse-grained to fine-grained and correlation information between modes.

[0073] Specifically, existing geological feature fusion technologies generally suffer from insufficient fusion depth and poor modality adaptability, failing to meet the needs of complex geological system analysis. On the one hand, traditional feature extraction methods (such as convolutional neural networks and single Transformer networks) often only extract geological features at a single level or in a single modality, failing to achieve progressive mining from shallow to deep layers. This results in extracted features that cannot fully reflect the complex nature of geological phenomena. For example, surface stratigraphic features extracted solely through a single network cannot reveal the potential impact of deep geological structures on disaster occurrence, leaving subsequent disaster risk assessments lacking crucial evidence. On the other hand, multi-source heterogeneous geological data contains various modalities with significant differences in structure, dimension, and semantic information. Traditional fusion methods (such as simple splicing and weighted summation) lack effective mining of inter-modal correlation information, easily leading to modal feature conflicts and the submergence of effective information, resulting in low-quality fused features and consequently affecting the accuracy of subsequent geological model construction. For instance, simply splicing historical structured stratigraphic data with real-time image-based monitoring data fails to reflect the spatiotemporal correlation between the two, significantly diminishing the fusion effect. This step takes a multi-source heterogeneous geological feature set as input and uses a cascaded Transformer with multiple sub-networks to first extract coarse-grained shallow features, then progressively mines fine-grained deep correlations. Each level embeds cross-modal attention fusion to generate a deep-level fused feature field that reflects the complex characteristics of geology. Leveraging the powerful modeling capabilities of the cascaded Transformer, it achieves a leap from shallow correlations to deep semantic fusion of multi-source heterogeneous geological data, providing a highly information-dense and highly complex unified feature representation for subsequent quantitative calculations and decision-making.

[0074] S203. Based on deep-level fusion of feature fields, incorporating geological prior knowledge rules to perform multi-objective collaborative optimization calculations, generating a risk-resource integrated three-dimensional geological model that quantifies the probability of disaster risk and the economic value of resources.

[0075] Geological a priori knowledge rules can be a set of deterministic or statistical rules formed based on long-term geological research, engineering practice experience, and industry standards, reflecting the evolution of geological phenomena, the interaction of geological factors, the mechanism of disaster occurrence, and the distribution patterns of resources. A risk-resource integrated three-dimensional geological model can be constructed through multi-objective collaborative optimization calculations, based on a deep integration of characteristic fields and geological a priori knowledge rules. It can intuitively present the geological structure, stratigraphic distribution, disaster risk probability distribution, and resource economic value distribution of the monitoring area in three-dimensional space.

[0076] Specifically, traditional geological model construction commonly suffers from prominent problems such as "single-objective orientation" and "lack of prior knowledge guidance," severely impacting the model's practicality and reliability and failing to meet the core requirement of modern geological engineering: "equal emphasis on safety and efficiency." Existing geological models often focus solely on either disaster risk assessment or resource value assessment, making it difficult to achieve simultaneous quantification and integrated presentation of both. For example, traditional disaster risk assessment models only consider the probability and impact range of geological disasters, neglecting the region's resource distribution and economic value. This can lead to excessive cost investment in disaster mitigation plans, resulting in over-management measures for high-value resource areas and resource waste. Conversely, traditional resource assessment models emphasize resource reserves and economic value, ignoring potential geological disaster risks during extraction and lacking targeted risk control considerations, easily leading to safety accidents, casualties, and significant economic losses. This step constructs a geological prior knowledge rule base, extracts core variables from a deep-fusion feature field, establishes a multi-objective optimization function with rule constraints, solves for quantified disaster risk and resource value, and combines 3D modeling technology to generate an integrated risk-resource 3D geological model by associating spatial coordinates. By combining prior geological knowledge with data-driven models, artificial intelligence is endowed with "geological common sense." Through multi-objective collaborative optimization, a three-dimensional integrated model that simultaneously quantifies and assesses safety risks and economic value 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.

[0077] S204. Based on the risk-resource integrated three-dimensional geological model, the decision mapping engine is used to optimize the targeted management of hidden disasters and the re-exploitation of resources, and generate a set of collaborative optimization engineering schemes.

[0078] A collaborative optimization engineering scheme set can be a collection of multiple collaborative optimization schemes for targeted governance of hidden disasters and resource re-exploitation.

[0079] Specifically, traditional engineering scheme development lacks intelligent and visual decision support tools. Schemes are often statically designed, unable to be adjusted in real-time according to dynamic changes in geological models, and difficult to simulate the implementation effects of different schemes. This results in schemes with poor adaptability and limited optimization space, failing to meet the dynamic needs of complex geological engineering. This step inputs 3D model data into a decision mapping engine, analyzes and identifies high-risk and high-value areas, simulates and evaluates multiple governance-mining scenarios, filters and optimizes non-compliant schemes, and finally organizes the implementation details of each scheme to form a collaborative optimization engineering scheme set. Advanced geological cognitive models are directly transformed into implementable engineering blueprints, achieving a closed-loop chain from "data -> model -> knowledge -> decision." Through automated and intelligent scheme generation, the efficiency and scientific nature of engineering planning are significantly improved, achieving deep synergy between disaster management and resource development in terms of space, time, and benefits. This provides a powerful intelligent decision-making tool for mine safety production, safe utilization of urban underground space, and site selection for major projects.

[0080] The method described in this embodiment first parallelly accesses historical paper data and real-time monitoring data streams, then uses NLP and signal processing techniques to perform structured analysis, spatiotemporal coding, and feature serialization to construct a multi-source heterogeneous geological feature set under a unified spatiotemporal framework. Subsequently, through a cascaded Transformer network and multi-level cross-modal attention computation, data-driven deep feature fusion from local details to global patterns is achieved, generating a deep-level fused feature field. Furthermore, formalized geological prior knowledge rules are used as constraints, and a multi-objective collaborative optimization algorithm drives the generation network to output an integrated 3D geological model that simultaneously quantifies disaster risk and resource value. Finally, based on this model, a decision mapping engine uses intelligent optimization algorithms to automatically search for and generate a set of Pareto-optimal engineering solutions that coordinate targeted management of hidden disasters and resource re-mining, and outputs this set to geological researchers. This method overcomes the long-standing problems of data silos and information dormancy in the geological field, generating a quantitative 3D model integrating risk and resources. It provides a unified scientific decision-making foundation for precise early warning and prevention of hidden disasters and the safe and green development of mineral resources, significantly improving the intelligence level and comprehensive benefits of geological engineering activities.

[0081] In some embodiments, the historical paper-based geological data set is subjected to structured parsing and encoding based on optical character recognition, visual models and spatiotemporal correlation to generate a structured mine history knowledge graph; multimodal feature extraction and spatiotemporal alignment processing are performed on real-time multi-source geological monitoring data streams to generate a unified benchmark multi-source observation spatiotemporal feature tensor; and a multi-source heterogeneous geological feature set is generated based on the entity attribute feature vectors extracted from the mine history knowledge graph and the multi-source observation spatiotemporal feature tensor.

[0082] Optical character recognition (OCR) is a technology that converts printed or handwritten text in images into machine-encodeable text. Visual models can be deep learning-based computer vision models, such as image segmentation models and object detection models. Structured parsing and encoding is the process of transforming unstructured raw data (text, drawings) into structured data with clear semantic relationships and spatiotemporal labels using technologies such as OCR and visual models. A mine history knowledge graph is a knowledge network that presents historical geological information of a mine 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 relationships between entities (such as the spatial relationship between strata and structures, and the attribution relationship between resource points and strata, etc.). It is the core result of structured parsing and encoding of historical paper-based geological data. Multimodal feature extraction is a processing technology for different types of data (modalities) in real-time multi-source geological monitoring data streams. Spatiotemporal alignment processing can be used because different monitoring devices (sources) may have different sampling frequencies, deployment locations, coordinate systems, and data formats. The purpose of spatiotemporal alignment processing is to unify these multi-source, heterogeneous real-time monitoring data onto the same spatiotemporal reference. The multi-source observation spatiotemporal feature tensor can be a multi-dimensional data array (tensor) formed by organizing all monitoring features under the same spatiotemporal unit (such as a certain time slice or a certain three-dimensional geological unit) after multimodal feature extraction and spatiotemporal alignment processing.

[0083] Specifically, traditional mine geological analysis has long faced the dilemma of data fragmentation: massive amounts of historical paper data (such as hand-drawn profiles and borehole logs) lie dormant in archives due to their unstructured nature, forming a "data graveyard" that cannot be directly understood by machines, and the mineralization laws and prior knowledge of disasters contained therein are difficult to reuse; while real-time monitoring data streams (such as microseismic and deformation data) are dynamic and rich, they lack a connection with historical background due to their multi-source heterogeneity and inconsistent spatiotemporal benchmarks, resulting in analysis remaining at isolated alarms and superficial correlations, unable to deeply reveal the causal mechanism between "current anomalies" and "historical structures". To address the above issues, this step first collects historical paper documents from the mine (such as exploration reports and borehole drawings). After high-definition scanning and restoration, optical character recognition (OCR) technology is used to extract textual values ​​(such as "drilling depth 150m, stratum lithology is sandstone"). A convolutional neural network (visual model) is used to parse the drawings, extract fault spatial coordinates, and then spatiotemporal encoding is added according to a unified standard (timestamp set to UTC2024-06-01T09:00:00, spatial coordinates adopt the National Geodetic Coordinate System 2000). A mine history knowledge graph is generated, from which stratum entity attribute feature vectors are extracted (such as "thickness 8m, compressive strength 30MPa"). At the same time, real-time monitoring data (such as displacement and audio) is received. First, a unified benchmark is established (time is unified in UTC, displacement unit is unified in millimeters, and spatial coordinates are the same as historical data). The mean displacement data (e.g., 0.3 mm / h) is extracted using the sliding window method. Crack features (e.g., width 2 mm) of borehole rock wall images are extracted using CNN. The peak value of the rock mass vibration audio spectrum (e.g., 50 Hz) is extracted using Fourier transform. Then, the 1-minute interval data is completed by linear interpolation, and the coordinates are calibrated by spatial interpolation (e.g., mapping equipment data with a deviation of 2 meters to grid nodes) to generate a multi-source observation spatiotemporal feature tensor. Finally, the formation entity vector is associated and fused with the real-time soil copper content data (e.g., 25 mg / kg) of the corresponding area using a spatiotemporal correlation algorithm. After removing redundant features, a multi-source heterogeneous geological feature set is generated.

[0084] The method provided in this embodiment generates a multi-source heterogeneous geological feature set by fusing entity attribute feature vectors with multi-source observation spatiotemporal feature tensors. This achieves deep correlation and complementarity between historical geological data and real-time monitoring data, so that the feature set contains both historical geological evolution laws and current geological dynamic changes, comprehensively reflecting the complex characteristics of the mining geological system.

[0085] In some embodiments, the historical paper-based geological data set includes old mine geological exploration report texts, old mine hand-drawn geological profile maps, old mine drilling logs, and historical mining records. Standardized geological entities and their semantic relationships are extracted from the old mine geological exploration report texts through cascaded processing of optical character recognition models and geological terminology. Linear symbols in the old mine hand-drawn geological profile maps are identified using a visual transformer and parsed and reconstructed into vectorized graphic objects with spatial coordinates and geological attributes. The structured text data from the old mine drilling logs and historical mining records is spatiotemporally aligned with the entities extracted from the old mine geological exploration report texts and old mine hand-drawn geological profile maps. Based on the semantic relationships, vectorized graphic objects, and aligned structured text data, a mine history knowledge graph is constructed.

[0086] Geological exploration reports for old mines can be written reports generated after geological exploration work is carried out in an old mining area. Hand-drawn geological profiles of old mines can be graphic data reflecting the spatial relationships of the old mine's vertical or horizontal geological structure, stratigraphic distribution, and ore body location, created by geological workers in historical periods through on-site surveying and combining exploration data. Drilling log sheets for old mines can be tabular paper records of real-time observations and descriptions of drilled core samples recorded by on-site technicians during drilling operations. Historical mining records can be ledgers or reports recording information such as stope location, mining volume, ore grade, and geological problems encountered (such as small faults and water seepage points) during past mine production activities. Optical character recognition (OCR) models can be computer vision models based on deep learning used to detect text regions in scanned document images and recognize them as computer-editable and processable coded text. Geological terminology can be a series of refined text processing procedures specifically for the geological field after general OCR has recognized the text. Geological semantic relationships can be logical connections with clear geological meanings connecting different geological entities; they are the edges connecting nodes in a knowledge graph. A visual transformer can be a deep learning model that applies the successful Transformer architecture from natural language processing to image recognition tasks. Linear symbols can be graphical markers representing different geological elements in hand-drawn geological profiles of old mines, such as solid lines representing defined stratigraphic boundaries, dashed lines representing inferred boundaries, sawtooth lines representing faults, and various dots, circles, and filled patterns representing different lithologies. Vectorized graphic objects can be the conversion of recognized linear symbols from raster image format to vector graphic format defined by mathematical equations (such as SVG paths).

[0087] Specifically, in traditional mine geological analysis, historical paper data (such as exploration report texts, hand-drawn maps, and logbooks) are fragmented, forming multiple information silos that cannot be communicated: the geological descriptions in the texts are isolated from the lines and symbols on the maps and the specific data in the tables. Comprehensive judgment relies on the manual memory, review, and subjective association of geological experts, which is extremely inefficient and prone to critical cognitive gaps due to personnel changes or omissions of details. Simply scanning and archiving the data only achieves digitization and is far from achieving knowledge-based understanding. Machines cannot understand that "the F5 fault mentioned in the report" is the same entity as "a sawtooth line on the profile" and "the breccia at a depth of 120 meters in the logbook". To address the above issues, this step first collects old mine geological exploration report texts, hand-drawn geological profile maps, drilling log tables, and historical mining records. After restoration, these are converted into electronic images and enhanced using a high-definition scanner (resolution set to 300 dpi). The report text is then input into an optical character recognition model to extract the original text. This is combined with a geological terminology database (containing terms such as "ore-bearing layer" and "conformable contact") for cascade processing to correct errors such as "sandstone" instead of "sandy rock." Entities such as "F2 normal fault" and "Jurassic sandstone strata" are extracted, along with the semantic relationship "F2 normal fault controls the western ore body." A visual transformer is used to analyze the profile maps, identifying 1:5000 scale scale and fault dashed line symbols, and reconstructing them into vectorized objects with coordinates. Structured data from drilling logs and mining records are extracted and correlated with the 2000 geodetic coordinate system of the mining area and the 1988 time base. Finally, a knowledge graph is constructed using a graph database, with nodes containing entity coordinates and other attributes, and edge tables showing semantic relationships.

[0088] By deeply linking text, drawings, and tabular data, this embodiment forms a unified and authoritative digital twin of the mine's geological conditions. This enables subsequent risk assessments, resource evaluations, and engineering planning to be based on a complete and consistent historical understanding, significantly improving the value of reusing old mine data and the level of intelligent management in the new era.

[0089] In some embodiments, the real-time multi-source geological monitoring data stream includes microseismic monitoring waveform time-series data, geochemical element sampling point data, and engineering drilling core logging data; source parameters are extracted from the microseismic monitoring waveform time-series data based on an event detection algorithm to form a microseismic event feature sequence with events as units; the geochemical element sampling point data is processed based on spatial interpolation and outlier analysis algorithms to generate a spatial distribution feature map of element concentration; the engineering drilling core logging data is analyzed based on natural language processing and lithology classification models to generate a lithology-structure discrete feature sequence with depth coordinates; the microseismic event feature sequence, element concentration spatial distribution feature map, and lithology-structure discrete feature sequence are uniformly transformed to the same three-dimensional geographic coordinate system and time reference to generate a multi-source observation spatiotemporal feature tensor.

[0090] Microseismic monitoring waveform time-series data can be data on the waveform signals generated by minute vibrations caused by rock fracturing, tectonic activity, etc., within the mining area, and their changes over time, acquired in real time by microseismic sensors. Geochemical element sampling point data can be data characterizing the content of various chemical elements (such as copper, lead, zinc, sulfur, and other elements related to mineral resources or the geological environment) in samples obtained from rock, soil, or groundwater samples collected at different spatial locations within the mining area, through laboratory testing or rapid on-site testing. Engineering drilling core logging data can be data formed by on-site observation, description, and recording of core samples obtained during engineering drilling operations in the mining area. This includes textual information such as core depth, lithological characteristics, tectonic phenomena, and mineral development, and is core data reflecting the true state of underground rock strata. Microseismic event feature sequences can be ordered sets of features formed by sorting multiple microseismic events extracted by event detection algorithms according to their occurrence time, with each event corresponding to a set of source parameters. Spatial interpolation and outlier analysis algorithms can be combined algorithms consisting of spatial interpolation algorithms (such as Kriging interpolation and inverse distance weighted interpolation) and outlier analysis algorithms (such as Grubbs test and box plot analysis). Element concentration spatial distribution feature maps can be images that, after processing with spatial interpolation algorithms, visually represent the concentration distribution of one or more chemical elements within a mining area (planar or three-dimensional space). Natural language processing and lithology classification models can be joint models composed of natural language processing (NLP) models (such as BERT and LSTM) and lithology classification models (such as ResNet and Support Vector Machines). Lithology-structure discrete feature sequences can be generated by parsing engineering drilling core logging data and arranging it in order of drilling depth, with each depth segment corresponding to a set of discrete features of "lithology category + structural features". A three-dimensional geographic coordinate system can be a three-dimensional coordinate system used to uniformly characterize the spatial location of a mining area, including longitude, latitude, and elevation (or depth), ensuring that various spatial data can be correlated under the same spatial reference.

[0091] Specifically, traditional methods for processing real-time multi-source geological monitoring data streams have significant drawbacks: microseismic monitoring waveform time-series data are simply filtered without using event detection algorithms to extract source parameters, making it impossible to quantify key information such as the depth and magnitude of microseismic events; geochemical element sampling point data are stored in discrete form without spatial interpolation and outlier analysis, making it difficult to present element distribution patterns and susceptible to outlier interference; engineering drilling core logging data relies on manual recording, resulting in non-standard lithological descriptions and a lack of depth correlation; more importantly, the spatiotemporal references for the three types of data are chaotic and cannot be integrated. To address these issues, real-time data is first collected: microseismic sensors are deployed at 500m intervals to acquire waveform data with a sampling frequency of 1000Hz; rock samples are collected in a 200m×200m grid, and element concentrations (e.g., copper 360mg / kg, lead 80mg / kg) are measured using a portable detector; during drilling, core data is recorded using a terminal equipped with a depth sensor (e.g., ZK203 well 320-325m: gray-black rock containing pyrite). The data were then processed separately: Microseismic data were processed using the STA / LTA event detection algorithm (short window 5s, long window 20s) to extract source parameters (e.g., depth 520m, magnitude ML1.2) and form an event sequence; elemental data were processed using the Grubbs test (significance 0.05) to remove outliers (e.g., zinc 1850mg / kg), and a 10m×10m concentration map was generated via Kriging interpolation; core data were processed using the BERT model to extract information, and the ResNet lithology classification model corrected "grayish-black rock" to basalt, forming a feature sequence based on depth. Finally, the national 2000 coordinate system and UTC time were unified (e.g., correcting the local microseismic time 13:10 to UTC 14:10), and integrated into a multi-source observation spatiotemporal feature tensor.

[0092] By unifying the three-dimensional geographic coordinate system and time reference, the spatiotemporal barriers of the three types of data are broken through the method provided in this embodiment. The generated multi-source observation spatiotemporal feature tensor can simultaneously carry the spatiotemporal correlation information of microseismic, elemental, and rock core data. This provides high-quality and highly correlated input data for subsequent cascaded Transformer networks to perform multi-level feature extraction and cross-modal fusion, significantly improving the accuracy of deep-level fusion feature fields.

[0093] In some embodiments, a multi-source heterogeneous geological feature set is input into multiple independent and structurally specialized cascaded Transformer networks. The cascaded Transformer networks include a single-source feature extraction layer, a shallow cross-modal interaction layer, and a deep cross-modal fusion layer connected in sequence. In the single-source feature extraction layer, independent Transformer encoders perform preliminary abstract feature analysis on entity attribute feature vectors from the mine history knowledge graph and different modal feature sequences from multi-source observation spatiotemporal feature tensors. In the shallow cross-modal interaction layer, a cross-modal attention mechanism is introduced to perform low-dimensional interaction and alignment on the preliminarily abstracted modal features. In the deep cross-modal fusion layer, a multi-head cross-attention mechanism is used to reorganize and deeply fuse high-dimensional semantic features, outputting a deep-level fused feature field that represents the geological spatiotemporal structure and attributes.

[0094] The single-source feature extraction layer can be the foundational layer of a cascaded Transformer network. Its core function is to independently abstract and extract features from the original features of a single modality, avoiding early interference between different modalities and laying the foundation for subsequent cross-modal interactions. The shallow cross-modal interaction layer can be an intermediate layer following the single-source feature extraction layer. Its core function is to achieve preliminary interaction, alignment, and association mining of features from different modalities in a low-dimensional feature space, addressing semantic differences and spatiotemporal misalignment issues among features from different modalities. The deep cross-modal fusion layer can be the core layer of a cascaded Transformer network, following the shallow interaction layer. Its core function is to deeply reorganize, strengthen associations, and integrate the pre-aligned cross-modal features in a high-dimensional semantic space, mining deep semantic connections between modalities. The cross-modal attention mechanism can be an attention calculation mechanism that calculates the association weights between features from different modalities, focuses on key association information, weakens irrelevant interference, and achieves feature alignment and interaction between modalities. It is the core technology of the shallow cross-modal interaction layer. The multi-head cross-attention mechanism is a mechanism that uses multiple parallel attention heads to calculate the cross-correlation weights between cross-modal features from different dimensions and perspectives, thereby realizing the deep reorganization and fusion of high-dimensional semantic features. It is the core technology of the deep cross-modal fusion layer.

[0095] Specifically, traditional multi-source geological data fusion techniques often employ simple feature stitching or unified processing using a single model, neglecting the inherent differences in physical meaning, scale gaps, and asymmetric interpretations between different modal data (such as descriptive historical knowledge, point-like microseismic events, continuous field element concentrations, and serialized lithological profiles). This results in a fusion process akin to mechanically mixing articles from different languages, where features cannot engage in effective semantic dialogue or mutual constraints. For instance, historical fault attributes cannot serve as prior background to guide and explain the spatiotemporal clustering patterns of microseismic events. Consequently, the fused feature representations are hybrid, poorly interpretable, and unable to support a profound and reliable joint inversion of complex subsurface geological spatiotemporal structures and attributes. To address the above issues, this step first acquires a multi-source heterogeneous geological feature set, including entity attribute feature vectors from a mine history knowledge graph, and modal features of multi-source observation spatiotemporal feature tensors (microseismic event feature sequences: parameters such as "depth 520m, magnitude ML1.2" sorted by time; elemental concentration spatial distribution feature map: copper element concentration map in a 10m×10m grid; lithology-structure discrete feature sequence: "300-305m: basalt, no fractures" depth sequence). Three cascaded Transformer networks with specialized structures are constructed: the temporal modal network is set to a 1000-step size to adapt to the microseismic sequence; the spatial modal network is augmented with spatial attention; and the entity modal network optimizes the semantic embedding layer. The single-source layer uses an independent Transformer encoder to mine the correlation between "F2 fault - eastern copper ore body" in the solid layer; the shallow layer introduces cross-modal attention to map features to the same low-dimensional space and calculate the weights of microseismic and lithological features to achieve alignment; the deep layer uses 8-head cross-attention to mine the correlation between "copper enrichment area - lithological fracturing - frequent microseismic events" from a causal perspective and reorganize high-dimensional semantic features; finally, it is organized according to the national 2000 three-dimensional coordinate system (latitude and longitude + depth) of the mining area to generate a deep-level fused feature field.

[0096] The method provided in this embodiment, through a carefully designed cascaded Transformer network structure, mimics the scientific cognitive process of geological experts who first analyze individual data, then establish preliminary correlations, and finally conduct in-depth synthesis. This makes the fusion of multi-source data no longer a black box operation, but a logically clear and step-by-step traceable intelligent processing flow, significantly improving the credibility and interpretability of the fusion results.

[0097] In some embodiments, for different modal data directly related to mineralization potential, Transformer encoders with different inductive biases are deployed to perform preliminary abstract feature screening for mineralization: a graph-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 mineralization-favorable lithological entities, thereby obtaining historical geological entity features; a time-aware Transformer encoder is used to process microseismic event feature sequences to extract their spatiotemporal aggregation and energy release trend features, thereby obtaining microseismic activity features; and a spatially-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, thereby obtaining geochemical anomaly features.

[0098] Preliminary abstract features can be obtained by initially screening and abstracting different modal data directly related to mineralization potential in the single-source feature extraction layer using a Transformer encoder with mineralization-oriented inductive bias, focusing on core information related to mineralization. Historical ore bodies can be mineral resource entities discovered and recorded during historical exploration and mining of the target mining area, possessing clear spatial location, scale, ore grade, and mineral type; they are the core historical evidence reflecting the mineralization potential of the mining area. Ore-controlling structures can be geological structural entities that control the formation, distribution, morphology, and scale of ore bodies. Mineralization-favorable lithological entities can be rock types within the mining area possessing physicochemical properties conducive to the formation and enrichment of mineral resources, such as sandstone, granite, and basalt with specific compositions. Historical geological entity features can be abstract features focusing on mineralization-related historical geological entities (historical ore bodies, ore-controlling structures, and mineralization-favorable lithologies) obtained after processing entity attribute feature vectors using a graph-enhanced Transformer encoder. Spatiotemporal clustering can be the concentrated distribution characteristics of microseismic events in the temporal and spatial dimensions. Energy release trend characteristics can be seen as the variation of energy released by microseismic events over time. Common trends include gradual energy increase, periodic energy fluctuations, and stable maintenance at a high energy level, which can indirectly reflect the activity intensity and stability of ore-controlling structures. Microseismic activity characteristics can be abstracted features obtained by processing microseismic event feature sequences through a time-series-aware Transformer encoder, which integrates the spatiotemporal clustering of microseismic events with energy release trend characteristics. Geochemical anomaly characteristics can be abstracted features obtained by processing elemental concentration spatial distribution feature maps through a spatial-aware Transformer encoder, focusing on the combined anomaly information of ore-forming elements (such as copper, lead, zinc, and sulfur).

[0099] Specifically, traditional single-source feature extraction uses a general-purpose Transformer encoder, which lacks a mineralization-oriented inductive bias and cannot specifically capture the core laws related to mineralization: when processing entity attribute feature vectors, it cannot encode the topological relationships of historical ore bodies, ore-controlling structures, etc.; when processing microseismic event feature sequences, it misses spatiotemporal aggregation and energy release trends; when processing element concentration maps, it is difficult to identify the spatial morphology of mineralization-related element combinations anomalies, and it also extracts a large amount of redundant information (such as ordinary rock strata properties and random microseismic noise), interfering with subsequent fusion. To address the above issues, this step first acquires the following ore-forming modal data: entity attribute feature vectors of the mine history knowledge graph (e.g., "F3 ore-controlling fault" strike N45°, dip angle 60°, "No. 5 ore body" length 800m, "Jurassic ore-forming sandstone" thickness 50m), microseismic event feature sequences (e.g., "depth 450m, magnitude ML1.3, coordinates E118°42′N30°26′" and "depth 460m, magnitude ML1.5, coordinates E118°43′N30°27′" ordered by time), and element concentration spatial distribution feature maps (copper, lead, and zinc 10m×10m grid maps, e.g., copper 350mg / kg and lead 280mg / kg in a certain area). Three types of encoders are then used for processing: the Transformer encoder with graph structure enhancement converts the entity into a graph (with "control relationship" edges set between the F3 fault and the No. 5 ore body), encoding the topological relationship to obtain historical geological entity features; the temporal-aware encoder uses a 12-hour sliding window to extract the spatiotemporal clustering of microseisms (such as the two events mentioned above being concentrated in the E118°42′-43′ region within 1 hour) and energy trends (magnitude increasing from 1.3 to 1.5), obtaining microseismic activity features; the spatial-aware encoder identifies the elemental anomaly banded morphology (such as the E118°40′-45′ region) and concentration gradient (copper at the center decreases from 800 mg / kg to 400 mg / kg 50m outward) according to the mineralization criteria (copper > 500 mg / kg and lead > 300 mg / kg), obtaining geochemical anomaly features; finally, the three types of features are standardized into a 256-dimensional vector for subsequent interaction.

[0100] By deploying specialized Transformer encoders with corresponding inductive biases for different data structures, the model architecture itself is made more in line with the inherent characteristics of various geological data. This enables deeper and more essential information mining in the basic feature extraction stage, avoiding information loss or distortion caused by a one-size-fits-all approach.

[0101] In some embodiments, a lightweight cross-modal attention module with spatial location as the key is introduced to perform initial cross-validation and localization focusing of mineral exploration clues: the preliminary abstract features of each mode are projected onto 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, historical geological entity features, microseismic activity features, and geochemical anomaly features are weighted and fused and information is complemented to generate anomaly-tectonic joint features that are initially coupled in space, so as to delineate favorable mineralization areas.

[0102] Lightweight cross-modal attention modules can be cross-modal attention computation units with simplified model parameters and computational complexity, using three-dimensional spatial coordinates as the core correlation basis. Cross-attention mechanisms, originating from data science, refer to using features of one modality (such as microseismic activity) to verify or interpret the reliability and geological significance of features of another modality (such as geochemical anomalies). Mutual attention weights can be calculated by the lightweight cross-modal attention module, representing the correlation importance (range 0-1) of different modal features at the same grid node; higher weights indicate stronger mineralization indicative significance of the modal feature for the current grid node. Anomaly-tectonic joint features can be coupled features that, after weighted fusion, simultaneously contain information on geochemical anomalies (material traces), historical geological structures (spatial basis), and microseismic activity (activation state), and can preliminarily characterize favorable mineralization conditions of "structure + anomaly + activation." Mineralization-favorable areas can be spatial regions within a mining area that simultaneously possess multiple favorable conditions such as mineralization spatial basis (structure), material traces (elemental anomalies), and activation signals (microseismic activity), and are key target areas for mineral resource exploration.

[0103] Specifically, traditional multimodal fusion methods often employ standard, global attention mechanisms or simple feature-level stitching during interaction, which has significant drawbacks in geological prospecting applications. The former has extremely high computational costs, making it difficult to handle massive amounts of 3D spatial data; the latter completely ignores the core geological constraint of the positional correspondence between different modal features in real 3D space, making it prone to generating false associations. For example, forcibly associating a deep microseismic event with a shallow and spatially unrelated geochemical anomaly leads to low credibility of clues and ambiguous target area location. To address the above issues, this step divides the mining area into a 10m×10m×5m grid according to the National 2000 coordinate system (e.g., node G001: E118°40′-40′10″, N30°25′-25′10″, depth 200-205m); historical geological features (e.g., the F3 ore-controlling fault) are projected onto the grid it passes through, along with microseismic activity characteristics (two microseismic events in area G001 within one hour, with magnitude ML1.1 increasing to 1.3) and geochemical anomaly characteristics (copper 620mg / kg, lead 350mg / kg in area G001). g, which meets the mineralization criteria, are projected onto the corresponding nodes; the lightweight cross-modal attention module is activated, and weights are assigned to G001 according to the mineralization logic (structure 0.4, microseismic 0.3, element 0.3). For nodes with only a single modal anomaly (such as G050 with only copper 550mg / kg), the element weight is reduced to 0.1; weighted fusion is performed (G001 fusion feature = structure × 0.4 + microseismic × 0.3 + element × 0.3) to supplement the missing information of a single mode; the fusion feature threshold is set to 0.6, and nodes with a threshold > 0.6 are selected to generate anomaly-structure joint features.

[0104] The method provided in this embodiment forces the features of different modalities to be aligned to a unified spatial grid and interacts based on local attention. This effectively realizes the mutual verification and supplementation of multi-source information in real geological space, filters out spatially inconsistent noise signals, and significantly improves the accuracy of identifying favorable mineralization locations.

[0105] In some embodiments, the anomaly-construction joint features coupled by spatial coupling are used as input, and the metallogenic pattern feature vector extracted from the mine history knowledge graph is introduced as a global query; through a multi-head cross-attention mechanism, the deep semantic matching degree between the metallogenic pattern feature vector and each spatial location is analyzed; the deep semantic matching degree is used as the fusion feature vector of each spatial location in the deep fusion feature field, and the fusion feature vector directly represents the similarity between this spatial location and the historical metallogenic pattern, that is, the potential probability of residual ore occurrence.

[0106] Metallogenic model feature vectors can be extracted from mine history knowledge graphs, representing vector data of "key metallogenic condition combinations," including historically verified core metallogenic elements (such as combinations of "ore-controlling faults + ore-forming element anomalies + tectonic activation signals"), and are a condensed expression of metallogenic regularities. Deep semantic matching degree can be calculated through a multi-head cross-attention mechanism, representing the degree of similarity (range 0-1) between the anomaly-tectonic joint features of a spatial location and the metallogenic model feature vector at a high-dimensional semantic level. A higher matching degree indicates that the metallogenic conditions at that location are closer to historical metallogenic models, and the probability of remaining mineralization is higher. The fused feature vector can be vector data integrating spatial location metallogenic information with deep semantic matching degree as its core, directly representing the similarity between that spatial location and historical metallogenic models; essentially, it is a quantification of the potential probability of remaining mineralization. The potential probability of remaining mineralization refers to the likelihood of unexploited remaining mineral resources existing at a certain spatial location, directly represented by the fused feature vector (deep semantic matching degree), and is a core quantitative indicator for resource potential assessment.

[0107] Specifically, while shallow interactions have achieved local spatial cross-validation of multi-source clues, they still cannot simulate the core cognitive process by which geological experts examine mineralization regularities from a global and systemic perspective. Traditional techniques either rely on subjective inferences based on expert experience or employ limited nonlinear fusion using conventional neural networks, both of which struggle to effectively capture the crucial "long-range dependency" relationships within mineralization systems—that is, the complex semantic relationships between different modal anomalies that may be spatially distant (e.g., hundreds of meters) but genetically closely related (e.g., deep rock masses and distant alteration zones). These global, high-level correlation patterns are key to accurately delineating hidden ore bodies and understanding the entire mineralization system. Without this deep fusion, predictions will remain at the level of mechanical superposition of local evidence, failing to generate highly reliable quantitative prediction maps that truly reflect the overall structure of the mineralization system. To address the above issues, this step first obtains the spatial coupling anomaly-construction joint features (e.g., grid node G001: F3 ore-controlling fault strike N45°, copper 620mg / kg, lead 350mg / kg, 2 ML1.2-ML1.3 micro-seismic events within 1 hour), then extracts the ore-forming pattern feature vector V from the mine history knowledge graph (e.g., "fault N30°-N60° + copper > 500mg / kg + micro-seismic event ≥ ML1.2", quantized as 256 dimensions); and configures a 6-head cross-attention module (each focusing on the fault strike). Using V as the global query and G001 as the key / value pair, the matching degree of each head is calculated (fault 0.9, element 0.85, microseismic 0.9, etc.), with an average of 0.88 as the deep semantic matching degree. This matching degree is integrated with the three-dimensional coordinates of G001 (E118°40′, N30°25′, 203m) to generate a fused feature vector, which is then embedded into the deep fused feature field according to the coordinates. The unstructured node G500 is verified, and its matching degree of 0.12 is determined to be a low residual ore probability zone to ensure the reliability of the features.

[0108] The method provided in this embodiment outputs a mineralization potential probability map that displays the spatial distribution of mineralization possibilities in an intuitive and quantitative form. It has high resolution and high reliability and can be directly used as the core scientific basis for selecting exploration target areas and deploying drilling projects, which greatly improves the accuracy and efficiency of mineral exploration.

[0109] In some embodiments, based on a deep-level fusion feature field, dual-path feature representations for disaster risk prediction and resource potential assessment are extracted in parallel. In the disaster risk quantification path, the deep-level fusion feature field is constrainedly fused with geomechanical a priori rules, and disaster risk probability values ​​are generated voxel by voxel through the coupled calculation of the activation degree of concealed structures and rock mass stability parameters. In the resource economic value quantification path, the fusion feature vector representing the matching degree of metallogenic model is coupled with ore grade, mining cost and market parameters through multi-objective optimization, and resource economic value index is generated voxel by voxel. The disaster risk probability values ​​and resource economic value index are integrated at the voxel level under a unified three-dimensional geological spatial framework to generate a risk-resource integrated three-dimensional geological model that includes both risk and value quantification attributes at each spatial location.

[0110] The disaster risk quantification path can be a processing flow focused on "calculating the probability of geological disasters occurring." Its core is to combine geomechanical principles to transform geological stability characteristics into quantifiable disaster risk indicators, encompassing the coupled calculation of concealed structural activation and rock mass stability parameters. A voxel can be the smallest volumetric unit in three-dimensional space with a uniform attribute value (such as density or probability), similar to a pixel in a two-dimensional image. Concealed structural activation can be a quantitative indicator (range 0-1) characterizing whether unexposed (concealed) geological structures (such as faults and fissures) in a mining area are active; the higher the activation, the greater the risk of landslides, collapses, and other disasters caused by tectonic activity. Rock mass stability parameters can be quantitative indicators characterizing the rock mass's resistance to deformation and damage, including rock mass integrity coefficient, compressive strength, and water content, directly affecting the likelihood of disasters occurring. The disaster risk probability value can be the probability of a geological disaster occurring at a certain voxel node (range 0-1, e.g., 0.8 represents high risk, 0.2 represents low risk) obtained after coupled calculation of concealed structural activation and rock mass stability parameters; it is the final output of the disaster risk quantification path. The path to quantifying the economic value of resources can focus on a process of "assessing the economic benefits of mineral resources." The core is to combine the matching degree of mineralization models, mining costs, and market conditions to transform the characteristics of mineralization value into quantifiable economic value indicators. The resource economic value index can be the quantified economic value of mineral resources in a region corresponding to a specific voxel node (e.g., 1200 yuan / cubic meter). A higher value indicates greater economic potential for the region's resources, and it is the final output of the resource economic value quantification path. Voxel-level integration can use voxel nodes in a unified three-dimensional framework as the smallest unit, binding the disaster risk probability value corresponding to each voxel to the resource economic value index, enabling each voxel to possess both "risk-value" attributes, achieving integrated association at the micro-level.

[0111] Specifically, traditional geological models separate disaster risk from resource value assessment, easily leading to decision-making contradictions such as "ignoring high-risk areas in high-value areas" or "over-managing low-value areas." Disaster quantification relies solely on single data such as microseismic frequency, lacking the constraints of prior geomechanical rules, resulting in results detached from the stress patterns of rock masses. Resource value assessment only considers ore grade, ignoring the matching degree of mineralization patterns (probability of remaining ore), easily leading to inflated values; and it can only integrate at the regional level, lacking sufficient accuracy. To address these issues, this step deeply integrates feature fields (voxel resolution 10m×10m×5m) and extracts geological stability features (e.g., voxel G001: F3 fault, 2 ML1.2-ML1.3 microseismic events per hour, basalt) and mineralization value features (G001 mineralization matching degree 0.88, copper grade 1.2%) in parallel. Disaster paths are integrated with geomechanical rules (e.g., activation degree = microseismic energy × rock mass fragmentation coefficient), calculating G001 activation degree as 0.3 and rock mass stability as 0.75, and obtaining the formula... The disaster risk probability is 0.28; the resource path is weighted (matching degree 0.4, grade 0.3, etc.) to calculate the G001 score of 0.892. Combined with the copper price of 60,000 yuan / ton, mining cost of 80 yuan / ton, and tax of 13%, the value index is 1260 yuan / cubic meter; the voxel risk is bound to the value (G001: 0.28+1260), and rendered by color mark (risk red = 0.7-1.0, value gold = 1000 yuan / cubic meter+), generating a clickable risk-resource integrated 3D geological model.

[0112] The method provided in this embodiment transforms continuous model probability outputs into target area maps with clear boundaries and explicit priority levels, enabling geological exploration personnel to directly apply them to engineering deployments without complex interpretation. This achieves seamless integration of artificial intelligence prediction and field practice, significantly improving the efficiency of exploration plan development.

[0113] Figure 3 This is a schematic diagram of the structure of a cascaded Transformer fusion system based on multi-source heterogeneous geological data provided in an embodiment of this application, as shown below. Figure 3 As shown, the cascaded Transformer fusion system 300 based on multi-source heterogeneous geological data in this 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;

[0114] The geological feature analysis module 301 is used to acquire historical paper-based geological data sets and real-time multi-source geological monitoring data streams, perform structured analysis and spatiotemporal encoding on the historical paper-based geological data sets, and extract and serialize features from 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 used to generate a deep-level fused feature field by performing multi-level feature extraction from coarse to fine and cross-modal attention fusion through a cascaded Transformer network based on the multi-source heterogeneous geological feature set. The 3D module construction module 303 is used to perform multi-objective collaborative optimization calculations based on the deep-level fused feature field and incorporating geological prior knowledge rules to generate a risk-resource integrated 3D geological model that quantifies the probability of disaster risk and the economic value of resources. The optimization scheme generation module 304 is used to optimize the targeted management of hidden disasters and the re-exploitation of resources through a decision mapping engine based on the risk-resource integrated 3D geological model to generate a set of collaborative optimization engineering schemes.

[0115] Optionally, the geological feature analysis module 301, when generating the multi-source heterogeneous geological feature set, is specifically used for: performing structured parsing and encoding on the historical paper geological data set based on optical character recognition, visual models, and spatiotemporal correlation to generate a structured mine history knowledge graph; performing multimodal feature extraction and spatiotemporal alignment processing on the real-time multi-source geological monitoring data stream to generate a unified benchmark multi-source observation spatiotemporal feature tensor; and generating the multi-source heterogeneous geological feature set based on the entity attribute feature vectors extracted from the mine history knowledge graph and the multi-source observation spatiotemporal feature tensor.

[0116] Optionally, the geological feature analysis module 301, when generating the structured mine history knowledge graph, is specifically used for: the historical paper geological data set including old mine geological exploration report text, old mine hand-drawn geological profile 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 cascade processing of optical character recognition model and geological professional terms; recognizing line symbols in the old mine hand-drawn geological profile map through visual transformer, and parsing and reconstructing them into vectorized graphic objects with spatial coordinates and geological attributes; aligning the structured text 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 profile map in a spatiotemporal manner; and constructing the mine history knowledge graph based on the geological semantic relationships, the vectorized graphic objects, and the aligned structured text data.

[0117] Optionally, the geological feature analysis module 301, when generating the multi-source observation spatiotemporal feature tensor based on the unified benchmark, is specifically used for: the real-time multi-source geological monitoring data stream including microseismic monitoring waveform time-series data, geochemical 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 with events as units; processing the geochemical element sampling point data based on spatial interpolation and outlier analysis algorithms to generate a spatial distribution feature map of element concentration; parsing the engineering drilling core logging data based on natural language processing and lithology classification models to generate a lithology-structure discrete feature sequence with depth coordinates; and uniformly converting 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 benchmark to generate the multi-source observation spatiotemporal feature tensor.

[0118] Optionally, when generating the deep-level fused feature field, the feature field generation module 302 is specifically used to: input the multi-source heterogeneous geological feature set into multiple independent and structurally specialized cascaded Transformer networks, wherein the cascaded Transformer networks include a single-source feature extraction layer, a shallow cross-modal interaction layer, and a deep cross-modal fusion layer connected in sequence; in the single-source feature extraction layer, an independent Transformer encoder performs preliminary abstract feature analysis on the entity attribute feature vectors from the mine history knowledge graph and the different modal feature sequences from the multi-source observation spatiotemporal feature tensors; in the shallow cross-modal interaction layer, a cross-modal attention mechanism is introduced to perform low-dimensional interaction and alignment on the preliminarily abstracted modal features; in the deep cross-modal fusion layer, a multi-head cross-attention mechanism is used to reorganize and deeply fuse high-dimensional semantic features, and output the deep-level fused feature field representing the geological spatiotemporal structure and attributes.

[0119] Optionally, the feature field generation module 302, based on the single-source feature extraction layer, is specifically used for: deploying Transformer encoders with different inductive biases for different modal data directly related to mineralization potential, and performing the preliminary abstract feature screening for mineralization: using a graph-enhanced Transformer encoder to process the entity attribute feature vector to strengthen the encoding of the topological relationships between historical ore bodies, ore-controlling structures, and mineralization-favorable lithological entities, and obtaining historical geological entity features; using a time-aware Transformer encoder to process the microseismic event feature sequence to extract its spatiotemporal aggregation and energy release trend features, and obtaining microseismic activity features; using a spatially-aware Transformer encoder to process the element concentration spatial distribution feature map to identify the spatial morphology and concentration gradient features of element combination anomalies, and obtaining geochemical anomaly features.

[0120] Optionally, the feature field generation module 302, when based on the shallow cross-modal interaction layer, is specifically used to: introduce a lightweight cross-modal attention module with spatial location as the key, and perform initial cross-validation and localization focusing of mineral exploration clues; project the preliminary abstract features of each modality onto 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 complementarity on the historical geological entity features, the microseismic activity features, and the geochemical anomaly features to generate spatially initially coupled anomaly-tectonic joint features to delineate favorable mineralization areas.

[0121] Optionally, the feature field generation module 302, when based on the deep cross-modal fusion layer, is specifically used to: take the spatially coupled anomaly-construction joint features as input, and introduce the metallogenic pattern feature vector extracted from the mine history knowledge graph as a global query; analyze the deep semantic matching degree of the metallogenic pattern feature vector and each spatial location through a multi-head cross-attention mechanism; and use the deep semantic matching degree as the fusion feature vector of each spatial location in the deep fusion feature field, wherein the fusion feature vector directly represents the similarity between this spatial location and the historical metallogenic pattern, i.e., the potential probability of residual ore occurrence.

[0122] Optionally, the three-dimensional module construction module 303 is specifically used for: extracting dual-path feature representations for disaster risk prediction and resource potential assessment in parallel based on the deep-level fusion feature field; in the disaster risk quantification path, constrainingly fusing the deep-level fusion feature field with geomechanical prior rules, and generating disaster risk probability values ​​on a voxel-by-voxel basis through coupled calculation of the activation degree of concealed structures and rock mass stability parameters; in the resource economic value quantification path, coupling the fusion feature vector representing the matching degree of mineralization model with ore grade, mining cost and market parameters through multi-objective optimization, and generating resource economic value indexes on a voxel-by-voxel basis; and integrating the disaster risk probability values ​​and the resource economic value indexes at the voxel level under a unified three-dimensional geological spatial framework to generate a risk-resource integrated three-dimensional geological model for each spatial location that simultaneously contains risk and value quantification attributes.

[0123] The system in this embodiment can be used to execute the methods of any of the above embodiments, and its implementation principle and technical effect are similar, so they will not be described again here.

Claims

1. A cascaded Transformer fusion method based on multi-source heterogeneous geological data, characterized in that, include: The system acquires historical paper-based geological data sets and real-time multi-source geological monitoring data streams, performs structured analysis and spatiotemporal coding on the historical paper-based geological data sets, and extracts and serializes features from the real-time multi-source geological monitoring data streams to generate a multi-source heterogeneous geological feature set. Real-time multi-source geological monitoring data stream is a continuous transmission stream of geological environment and geological state related data collected in real time by various geological monitoring equipment from different monitoring dimensions, including microseismic monitoring waveform time series data, geochemical element sampling point data, and engineering drilling core logging data; Multi-source heterogeneous geological feature sets integrate historical geological data features that have undergone structured analysis and spatiotemporal coding, as well as real-time geological monitoring data features that have undergone feature extraction and serialization. Based on the aforementioned multi-source heterogeneous geological feature set, a deep-level fused feature field is generated by performing multi-level feature extraction from coarse to fine and cross-modal attention fusion through a cascaded Transformer network. Based on the deep-level fusion feature field, multi-objective collaborative optimization calculations are performed by incorporating geological prior knowledge rules to generate a risk-resource integrated three-dimensional geological model that quantifies the probability of disaster risk and the economic value of resources. Based on the aforementioned risk-resource integrated three-dimensional geological model, a decision mapping engine is used to optimize the targeted management of hidden disasters and the re-exploitation of resources, generating a set of collaborative optimization engineering solutions. The generation of the deep fusion feature field includes: The multi-source heterogeneous geological feature set is input into multiple independent and structurally specialized cascaded Transformer networks. The cascaded Transformer network includes a single-source feature extraction layer, a shallow cross-modal interaction layer, and a deep cross-modal fusion layer connected in sequence. In the single-source feature extraction layer, an independent Transformer encoder is used to perform preliminary abstract feature analysis on the entity attribute feature vectors from the mine history knowledge graph and the different modal feature sequences from the spatiotemporal feature tensor of multi-source observations. In the shallow cross-modal interaction layer, a cross-modal attention mechanism is introduced to perform low-dimensional interaction and alignment of the modal features after preliminary abstraction. In the deep cross-modal fusion layer, high-dimensional semantic features are reorganized and deeply fused through a multi-head cross-attention mechanism, and the deep fusion feature field that characterizes the geological spatiotemporal structure and attributes is output. The single-source feature extraction layer includes: For different modal data directly related to mineralization potential, Transformer encoders with different inductive biases are deployed to perform the preliminary abstract feature screening for mineralization: The entity attribute feature vectors are processed using a graph-enhanced Transformer encoder to strengthen the encoding of topological relationships between historical ore bodies, ore-controlling structures, and mineralization-favorable lithological entities, thereby obtaining historical geological entity features. A time-aware Transformer encoder is used to process the microseismic event feature sequence to extract its spatiotemporal clustering and energy release trend features, thereby obtaining the microseismic activity features. The spatially aware Transformer encoder is used to process the spatial distribution feature map of element concentration in order to identify the spatial morphology and concentration gradient features of element combination anomalies and obtain geochemical anomaly features.

2. The method according to claim 1, characterized in that, The generation of the multi-source heterogeneous geological feature set includes: The historical paper-based geological data set is subjected to structured analysis and encoding based on optical character recognition, visual models and spatiotemporal correlation to generate a structured mine history knowledge graph; Multimodal feature extraction and spatiotemporal alignment processing are performed on the real-time multi-source geological monitoring data stream to generate a unified benchmark multi-source observation spatiotemporal feature tensor; The multi-source heterogeneous geological feature set is generated based on the entity attribute feature vectors extracted from the mine history knowledge graph and the multi-source observation spatiotemporal feature tensor.

3. The method according to claim 2, characterized in that, The generation of the structured mine history knowledge graph includes: The historical paper-based geological data collection includes texts of geological exploration reports of old mines, hand-drawn geological profiles of old mines, drilling log tables of old mines, and historical mining records. By cascading the optical character recognition model with geological terminology, standardized geological entities and their geological semantic relationships are extracted from the old mine geological exploration report text. The line symbols in the hand-drawn geological profile of the old mine are identified by a visual transformer, and then parsed and reconstructed into vectorized graphic objects with spatial coordinates and geological attributes. The structured text data in the old mine drilling log table and the historical mining record table are spatiotemporally aligned with the entities extracted from the old mine geological exploration report text and the old mine hand-drawn geological profile map. Based on the geological semantic relationships, the vectorized graphical objects, and the aligned textual structured data, the mine history knowledge graph is constructed.

4. The method according to claim 2, characterized in that, The multi-source observation spatiotemporal feature tensor for generating a unified benchmark includes: The real-time multi-source geological monitoring data stream includes microseismic monitoring waveform time-series data, geochemical element sampling point data, and engineering drilling core logging data. Based on the event detection algorithm, source parameters are extracted from the time-series data of the microseismic monitoring waveforms to form a microseismic event feature sequence with events as units; The geochemical element sampling point data are processed based on spatial interpolation and outlier analysis algorithms to generate a spatial distribution feature map of element concentration. The engineering drilling core logging data is analyzed based on natural language processing and lithology classification model to generate a lithology-structure discrete feature sequence with depth coordinates. The microseismic event feature sequence, the elemental concentration spatial distribution feature map, and the lithology-structure discrete feature sequence are uniformly transformed to the same three-dimensional geographic coordinate system and time reference to generate the multi-source observation spatiotemporal feature tensor.

5. The method according to claim 4, characterized in that, The shallow cross-modal interaction layer includes: A lightweight cross-modal attention module with spatial location as the key is introduced to perform initial cross-validation and localization focusing of mineral exploration clues: The preliminary abstract features of each modality are projected onto a unified gridded space based on their three-dimensional spatial coordinates, and the mutual attention weights between features of different modalities are analyzed through a lightweight cross-attention mechanism. Based on the aforementioned mutual attention weights, the historical geological entity characteristics, the microseismic activity characteristics, and the geochemical anomaly characteristics are weighted, fused, and complemented to generate spatially preliminarily coupled anomaly-tectonic joint features, thereby delineating favorable mineralization areas.

6. The method according to claim 5, characterized in that, The deep cross-modal fusion layer includes: Using the spatially coupled anomaly-structure joint features as input, the metallogenic pattern feature vector extracted from the mine history knowledge graph is introduced as a global query. The deep semantic matching degree between the feature vector of the mineralization mode and each of the spatial locations is analyzed by using a multi-head cross-attention mechanism. The deep semantic matching degree is used as the fusion feature vector of each spatial location in the deep fusion feature field. The fusion feature vector directly represents the similarity between this spatial location and the historical mineralization pattern, that is, the potential probability of residual mineralization.

7. The method according to claim 6, characterized in that, The risk-resource integrated three-dimensional geological model that generates quantitative disaster risk probability and resource economic value includes: Based on the deep fusion feature field, dual-path feature representations for disaster risk prediction and resource potential assessment are extracted in parallel; In the disaster risk quantification path, the deep-level fusion feature field is constrainedly fused with the geomechanical a priori rules, and disaster risk probability values ​​are generated voxel by voxel through the coupling calculation of the activation degree of concealed structures and rock mass stability parameters. In the path of quantifying the economic value of resources, the fusion feature vector representing the matching degree of mineralization model is coupled with ore grade, mining cost and market parameters through multi-objective optimization, and the resource economic value index is generated on a voxel-by-voxel basis. The disaster risk probability value and the resource economic value index are integrated at the voxel level within a unified three-dimensional geological spatial framework to generate a risk-resource integrated three-dimensional geological model for each spatial location that simultaneously includes risk and value quantification attributes.

8. A cascaded Transformer fusion system based on multi-source heterogeneous geological data, characterized in that, Applied to the method as described in any one of claims 1-7, comprising: The geological feature analysis module is used to acquire historical paper-based geological data sets and real-time multi-source geological monitoring data streams, perform structured parsing and spatiotemporal encoding on the historical paper-based geological data sets, and extract and serialize features from the real-time multi-source geological monitoring data streams to generate multi-source heterogeneous geological feature sets. The feature field generation module is used to generate a deep-level fused feature field by performing multi-level feature extraction from coarse to fine and cross-modal attention fusion through a cascaded Transformer network based on the multi-source heterogeneous geological feature set. The three-dimensional module construction module is used to perform multi-objective collaborative optimization calculations based on the deep-level fusion feature field and incorporate geological prior knowledge rules to generate a risk-resource integrated three-dimensional geological model that quantifies the probability of disaster risk and the economic value of resources. The optimization scheme generation module is used to optimize the targeted management of hidden disasters and the re-exploitation of resources based on the risk-resource integrated three-dimensional geological model through a decision mapping engine, and generate a set of collaborative optimization engineering schemes.

Images