A mineral resource dynamic prediction and mining management system

Abstract

The present application relates to the technical field of mineral resources management, and is especially a mineral resources dynamic prediction and mining management system, comprising a data sensing and fusion layer, a unified digital twin model, a dynamic prediction and decision-making agent, and a visualization and interactive control layer; the data sensing and fusion layer collects structured data such as geological exploration and mining environment and unstructured market data, and generates a unified space-time tensor after processing; the unified digital twin model generates a dynamic comprehensive mine situation map containing resource reserve risk economic indicators through a three-dimensional convolutional neural network embedded with an attention mechanism; the dynamic prediction and decision-making agent predicts future reserves and geological risks, constructs a double-objective optimization model to generate an optimal mining path equipment scheduling and resource allocation scheme; and the visualization and interactive control layer presents the mine state and decision-making scheme in three dimensions and provides an interactive interface. The present application improves data utilization and decision-making scientificity, reduces safety risks, and promotes the intelligentization of mine management.

Description

Technical Field

[0001] This invention relates to the field of mineral resource management technology, specifically to a dynamic prediction and mining management system for mineral resources. Background Technology

[0002] The current field of mineral resource management faces several technical limitations. Firstly, there is insufficient capacity for multi-source data integration. Structured data from geological exploration, mining operations, and environmental monitoring are often isolated from unstructured market data such as industry reports, commodity transactions, and policy announcements, lacking effective fusion methods. This results in significant data heterogeneity and temporal discrepancies, making it difficult to obtain comprehensive and accurate data support for decision-making. Secondly, the application of digital twins is incomplete. Existing models struggle to integrate multimodal data such as 3D geological information, equipment operating status, environmental indicators, and market dynamics in real time. They cannot dynamically extract spatiotemporal fusion features, resulting in outdated mining area status maps that cannot accurately synchronize with the resources, risks, and economic conditions of the actual mining area. The system suffers from several drawbacks: firstly, it lacks a multi-dimensional understanding of the situation; secondly, its dynamic prediction and decision-making capabilities are limited, with reserve predictions often neglecting the correlation with geological risks, and mining scheme optimization frequently focusing on a single objective, making it difficult to simultaneously maximize net present value and minimize mining safety risks, thus failing to flexibly adapt to real-time changes in the mining area; and thirdly, its visualization and interactive functions are weak, with data often displayed in a two-dimensional format, resulting in an unintuitive presentation, rudimentary interactive interfaces, and difficulty in quickly linking the system to re-optimize after parameter adjustments, affecting decision-making efficiency and accuracy, and failing to meet the actual needs of modern intelligent mine management. Therefore, this paper proposes a dynamic prediction and mining management system for mineral resources to address these issues. Summary of the Invention

[0003] The purpose of this invention is to provide a dynamic prediction and mining management system for mineral resources to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A dynamic prediction and mining management system for mineral resources, comprising:

[0006] The data perception and fusion layer collects structured geological exploration data, mining operation data, environmental monitoring data, and unstructured market dynamic data in real time through a sensor network deployed at the mine site and external data interfaces; it uses an entity relationship extraction model based on natural language processing to structure the unstructured market dynamic data and generate market intelligence vectors.

[0007] The unified digital twin model is used to receive and process multimodal data. The unified digital twin model uses a three-dimensional convolutional neural network with an embedded attention mechanism to fuse the spatiotemporal features of geological bodies, mining equipment, environmental indicators and market intelligence vectors, and outputs a dynamically updated comprehensive mining area situation map that includes resource reserves, grade distribution, geological risk factors and economic indicators.

[0008] A dynamic prediction and decision-making intelligent agent is used for collaborative decision-making based on a comprehensive mining area situation map; the decision-making intelligent agent includes:

[0009] The reserve-risk prediction sub-agent uses a gated recurrent unit network to learn the spatiotemporal sequence data in the comprehensive mining area situation map, and dynamically outputs the probability distribution of resource reserves and geological risk level of different zones in a specific future period.

[0010] The mining-economic optimization sub-agent receives the output of the reserve-risk prediction sub-agent at its input end and constructs a mixed integer programming model with the dual objectives of maximizing net present value and minimizing mining safety accident risk. The mixed integer programming model takes the probability distribution of resource reserves and the geological risk level as key constraints, automatically solves and generates the optimal mining path planning, equipment scheduling scheme and resource allocation strategy.

[0011] The visualization and interactive control layer connects to the dynamic prediction and decision-making intelligent agent. It is used to render the unified digital twin model and the optimal mining path planning in a three-dimensional visualization form, and to provide users with visualized early warning signals and decision parameter adjustment interfaces based on geological risk level and mining safety accident risk.

[0012] As a preferred approach, the processing of unstructured market dynamics data and the fusion of multi-source data in the data perception and fusion layer include:

[0013] Unstructured text data streams are obtained in real time from industry reports, commodity trading platforms and policy announcements through distributed web crawlers, and the obtained text data is preprocessed by denoising and standardization.

[0014] A pre-trained natural language processing model is used to perform entity recognition and relation extraction on the pre-processed text, extracting key entities and their semantic relationships, including mineral resource categories, price fluctuation trends, supply and demand relationships, and policy orientations.

[0015] A dynamic knowledge graph is constructed based on the extracted entity relationships, and the semantic information in the knowledge graph is mapped into a low-dimensional dense market intelligence vector through a graph embedding algorithm.

[0016] The market intelligence vector is aligned with geological exploration data and mining operation data collected by sensor networks in a multimodal manner, and a cross-modal attention mechanism is used to correct timestamps and compensate features for data with inconsistent time series.

[0017] The corrected multi-source data is encoded into a spatiotemporal tensor sequence that can be received by the unified digital twin model through the feature-level fusion module. Each spatiotemporal unit contains a fusion feature representation of geological attributes, equipment status, environmental indicators, and market influencing factors.

[0018] As a preferred solution, the unified digital twin model receives and processes multimodal data to achieve spatiotemporal feature fusion and dynamic generation of a comprehensive mining area situation map, including:

[0019] It receives multimodal data from the data perception and fusion layer, including three-dimensional voxel data of geological bodies, time series of mining equipment operation status, point cloud data of environmental indicator monitoring, and market intelligence vectors. It performs preliminary spatiotemporal registration on the multimodal data and maps data from different sources onto a unified three-dimensional spatial coordinate system and time axis of the mining area.

[0020] Multi-resolution fusion is performed on the spatiotemporally registered multimodal data. The sampling frequency and spatial granularity of geological bodies, equipment status, environmental indicators and market intelligence vectors are aligned by a feature completion method based on spatiotemporal interpolation to generate a consistent multimodal spatiotemporal tensor.

[0021] A three-dimensional convolutional neural network with an attention mechanism is embedded with a multimodal spatiotemporal tensor input. The attention module dynamically calculates the weight distribution of different modal features in the spatiotemporal dimension to highlight the impact of geological anomalies, changes in key equipment status, and market fluctuations. The fused spatiotemporal feature representation is extracted layer by layer through three-dimensional convolution operations.

[0022] The output feature map of a three-dimensional convolutional neural network is decoded into a multi-dimensional situation matrix containing the probability distribution of resource reserves, grade gradient changes, geological risk factors and economic indicators through a fully connected layer. The comprehensive mining area situation map is dynamically updated based on a time-series sliding window mechanism to achieve real-time synchronous rendering of the mining area status.

[0023] As a preferred approach, the reserve-risk prediction sub-agent uses a gated recurrent unit network to learn from the spatiotemporal sequence data in the comprehensive mining area situation map, dynamically outputting the probability distribution of resource reserves and geological risk levels in different zones within a specific future period, including:

[0024] Multidimensional spatiotemporal sequence data is extracted from the comprehensive mining area situation map generated by the unified digital twin model. The multidimensional spatiotemporal sequence data includes the temporal changes of resource reserves, grade distribution, geological risk factors and economic indicators.

[0025] The extracted multidimensional spatiotemporal sequence data is preprocessed with temporal alignment and normalization to eliminate dimensional differences and ensure data consistency, thereby generating a standardized spatiotemporal sequence.

[0026] The standardized spatiotemporal sequence is input into the gated recurrent unit network, and the temporal dependencies are learned through the gating mechanism, and the hidden state sequence is output.

[0027] Based on the hidden state sequence, a probability distribution of resource reserves in different partitions within a specific future period is generated through a fully connected layer and a probability output layer. The probability output layer uses a softmax function to represent the probability of different reserve intervals.

[0028] Meanwhile, based on the hidden state sequence, a geological risk level is generated through a risk classification layer, where the risk classification layer maps the hidden state to a predefined risk category;

[0029] The probability distribution of resource reserves and geological risk levels are dynamically updated to reflect real-time data changes and provide input for the mining-economic optimization sub-agent.

[0030] As a preferred approach, the input of the mining-economic optimization sub-agent is coupled with the output of the reserve-risk prediction sub-agent, and a mixed-integer programming model is constructed with the dual objectives of maximizing net present value and minimizing mining safety accident risk. The mixed-integer programming model uses the probability distribution of resource reserves and the geological risk level as key constraints, automatically solving and generating optimal mining path planning, equipment scheduling schemes, and resource allocation strategies, including:

[0031] Based on the probability distribution of resource reserves, quantile estimates of resource reserves are calculated using probabilistic statistical methods, and these quantile estimates are used as the resource reserve constraint boundary of the mixed integer programming model to reflect the impact of resource uncertainty.

[0032] Based on geological risk levels, risk levels are transformed into corresponding risk probability values ​​through a risk mapping function, and the risk probability values ​​are used as safety risk constraints in a mixed-integer programming model to quantify potential risks in the mining process.

[0033] A net present value (NPV) maximization objective function is constructed, where NPV calculation is based on resource extraction volume, market price forecasts, discount rates, equipment operating costs, and environmental compliance costs, and market dynamic data is integrated to enhance the real-time nature of economic assessment.

[0034] An objective function for minimizing mining safety accident risks is constructed, in which risk calculation is based on geological risk probability values, historical equipment failure data, human operation error coefficients, and environmental monitoring indicators to comprehensively cover safety factors;

[0035] The bi-objective function is transformed into a solvable single-objective optimization problem through a multi-objective optimization algorithm. The Pareto front generation method is used to generate a set of non-dominated solutions, and the optimal candidate solution is selected from the non-dominated solutions based on the decision-maker's preference weights.

[0036] The optimal candidate solution is verified for feasibility and locally optimized using a mixed-integer programming solver. The mining sequence and equipment allocation are adjusted to meet all constraints, and the final optimization result is output.

[0037] Based on the final optimization results, a detailed mining path planning sequence, equipment scheduling schedule, and resource allocation strategy are generated, and the strategy is dynamically updated to respond to real-time data changes, ensuring the adaptability and efficiency of the mining process.

[0038] As a preferred solution, the visualization and interactive control layer is used to render the unified digital twin model and optimal mining path planning in a 3D visualization format, and provides users with visualized early warning signals and decision parameter adjustment interfaces based on geological risk levels and mining safety accident risks, including:

[0039] It receives output data from the dynamic prediction and decision-making intelligent agent, including comprehensive mining area situation map, resource reserve probability distribution, geological risk level, optimal mining path planning and equipment scheduling scheme, and parses and converts the output data to generate structured scene data suitable for 3D rendering engine.

[0040] Based on structured scene data, a dynamic 3D mining scene is constructed through a real-time graphics rendering pipeline. The geological body uses voxel rendering technology to represent the distribution of resource grades with color gradients. The mining path planning is displayed by overlaying dynamic highlighted path lines and equipment animation models. Environmental indicators are generated by fusion of point cloud data to form a thermal layer.

[0041] An integrated risk warning engine calculates the comprehensive risk value of each spatial unit based on the parsed geological risk level and mining safety accident risk data through a risk mapping algorithm. The warning signal is then rendered in a 3D scene using color coding, flashing markers, and risk isosurfaces, with high-risk areas highlighted by a red gradient.

[0042] Deploy an interactive control panel to provide users with a multi-dimensional decision parameter adjustment interface, including a risk tolerance slider, an economic weight input box, and a mining priority selector. After users modify the parameters in real time through the graphical interface, the system automatically encapsulates the adjusted parameters as constraints and feeds them back to the dynamic prediction and decision-making agent for re-optimization.

[0043] Based on the re-optimization results triggered by user parameter adjustments, the mining path planning, equipment scheduling scheme and early warning signals in the 3D scene are dynamically refreshed through an incremental update mechanism to ensure that the visualization content is synchronized with the latest decision data;

[0044] It supports a multi-view collaborative visualization mode, allowing users to simultaneously browse 3D scenes, 2D planar plots, and time-series risk curves. Through view linkage technology, it enables data drill-down and interactive exploration from any perspective to assist users in comprehensive decision analysis.

[0045] As can be seen from the technical solution provided by the present invention above, the beneficial effects of the mineral resource dynamic prediction and mining management system provided by the present invention are:

[0046] Multi-source data fusion processing is more accurate and efficient. It can collect structured data such as geological exploration and mining operation environment monitoring in real time through sensor networks, and obtain unstructured market data such as industry reports and commodity trading policies through distributed web crawlers. Then, it uses natural language processing to extract entity relationships and construct a dynamic knowledge graph, transforming unstructured data into low-dimensional market intelligence vectors. With the help of cross-modal attention mechanism to correct inconsistent data in time series, it finally generates a unified spatiotemporal tensor sequence, completely eliminating data heterogeneity and time series bias, avoiding decision bias caused by isolated or inaccurate data, and providing comprehensive and high-quality data support for subsequent mining area analysis.

[0047] A unified digital twin model enables real-time and accurate mapping of mining area status. By embedding an attention mechanism into a 3D convolutional neural network, it performs spatiotemporal feature fusion on the point cloud of the geological body's 3D voxel equipment runtime environment and market intelligence vector. This dynamically highlights the key status changes of equipment in geologically abnormal areas and the impact of market fluctuations. Ultimately, it generates a comprehensive mining area status map that includes the probability distribution of resource reserves, grade gradient changes, geological risk factors, and economic indicators. Furthermore, it is updated in real time based on a time-series sliding window to ensure that the status map is completely synchronized with the actual mining area status. This allows managers to grasp the overall dynamics of the mining area in real time without on-site inspections, avoiding reliance on lagging data for decision-making.

[0048] The dynamic prediction and decision-making agent enhances the scientific rigor and practicality of decision-making. The reserve-risk prediction sub-agent learns from the spatiotemporal sequence data of the mining area using a gated cyclic unit network, and can accurately output the probability distribution of resource reserves and geological risk levels of different zones in a specific future period, providing accurate predictions for subsequent optimization. The mining-economic optimization sub-agent uses this prediction as a constraint to construct a dual-objective mixed integer programming model that maximizes net present value and minimizes mining safety accident risks. It generates non-dominated solutions through the Pareto front and selects the optimal solution by combining decision-maker preferences. The generated mining path plan can avoid high-risk areas, the equipment scheduling scheme can match the mining progress, and the resource allocation strategy can balance the needs of each link, effectively balancing mining economy and safety, and avoiding resource waste or safety accidents caused by blind mining.

[0049] The visualization and interactive control layer lowers the decision-making threshold and improves interaction efficiency. Through voxel rendering, the geological resource grade is presented with color gradients, mining plans are displayed with dynamically highlighted path lines, and environmental thermal layers are generated by point cloud fusion. Color-coded flashing risk isosurfaces highlight risk areas, transforming complex mining data into intuitive 3D scenes. At the same time, interactive interfaces such as risk tolerance sliders and economic weight input boxes are deployed. After users adjust parameters, the system automatically feeds back to the decision-making agent for re-optimization. Incremental updates only refresh the changed parts, eliminating the need for a full scene redraw, which greatly reduces the difficulty of understanding decision information and reduces the waiting time after parameter adjustments, helping managers make scientific decisions quickly.

[0050] The system is highly adaptable and can continuously optimize the mining process. It can receive real-time updates from sensors in the mining area and market dynamics, and synchronously adjust the data fusion results, digital twin potential map, prediction results, and decision-making schemes. For example, when the actual mining volume in a certain area exceeds the prediction, it can promptly lower the reserve constraint for that area and optimize subsequent mining plans. When market prices fluctuate, it can update the net present value calculation parameters to adjust mining priorities, effectively avoiding resource misallocation or safety hazards caused by data lag. In the long term, it can continuously improve the utilization rate of mine resources, reduce operating costs, and reduce the incidence of safety accidents, promoting the transformation of mine management towards intelligence and efficiency. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the structure of a dynamic prediction and mining management system for mineral resources according to the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0054] like Figure 1 As shown, this embodiment of the invention provides a dynamic prediction and mining management system for mineral resources, including:

[0055] The data perception and fusion layer collects structured geological exploration data, mining operation data, environmental monitoring data, and unstructured market dynamic data in real time through a sensor network deployed at the mine site and external data interfaces; it uses an entity relationship extraction model based on natural language processing to structure the unstructured market dynamic data and generate market intelligence vectors.

[0056] The unified digital twin model is used to receive and process multimodal data. The unified digital twin model uses a three-dimensional convolutional neural network with an embedded attention mechanism to fuse the spatiotemporal features of geological bodies, mining equipment, environmental indicators and market intelligence vectors, and outputs a dynamically updated comprehensive mining area situation map that includes resource reserves, grade distribution, geological risk factors and economic indicators.

[0057] A dynamic prediction and decision-making intelligent agent is used for collaborative decision-making based on a comprehensive mining area situation map; the decision-making intelligent agent includes:

[0058] The reserve-risk prediction sub-agent uses a gated recurrent unit network to learn the spatiotemporal sequence data in the comprehensive mining area situation map, and dynamically outputs the probability distribution of resource reserves and geological risk level of different zones in a specific future period.

[0059] The mining-economic optimization sub-agent receives the output of the reserve-risk prediction sub-agent at its input end and constructs a mixed integer programming model with the dual objectives of maximizing net present value and minimizing mining safety accident risk. The mixed integer programming model takes the probability distribution of resource reserves and the geological risk level as key constraints, automatically solves and generates the optimal mining path planning, equipment scheduling scheme and resource allocation strategy.

[0060] The visualization and interactive control layer connects to the dynamic prediction and decision-making intelligent agent. It is used to render the unified digital twin model and the optimal mining path planning in a three-dimensional visualization form, and to provide users with visualized early warning signals and decision parameter adjustment interfaces based on geological risk level and mining safety accident risk.

[0061] In this embodiment, the data perception and fusion layer is the core of the data input for the dynamic prediction and mining management system of mineral resources. It collects multi-type data through multiple channels, and after professional processing and deep fusion, it transforms the scattered and heterogeneous data into a unified and standardized spatiotemporal tensor sequence, providing high-quality data support for the subsequent unified digital twin model and ensuring the accuracy and real-time performance of the system's overall decision-making.

[0062] The core responsibility of the data perception and fusion layer is to achieve the collection, processing, alignment, and fusion of data across all dimensions. On the one hand, it acquires structured data such as geological exploration, mining operations, and environmental monitoring, as well as unstructured data related to market dynamics, in real time through field sensor networks and external data interfaces. On the other hand, it performs structured transformation on unstructured data, fusing multi-source data into a unified spatiotemporal tensor sequence through multimodal alignment, temporal correction, and other operations. This ensures the consistency, integrity, and timeliness of the data, laying the foundation for subsequent spatiotemporal feature fusion and decision analysis in models. The data perception and fusion layer includes:

[0063] Multi-source data acquisition unit:

[0064] On-site structured data acquisition: Through a sensor network deployed at the mine site, geological exploration data, including stratigraphic structure, rock type, and resource grade, is collected in real time; mining operation data, including mining progress, equipment operating parameters, and operation location, is collected; and environmental monitoring data, including air quality, soil moisture, and hydrological indicators, is collected. All data is transmitted to the system in real time to ensure data immediacy.

[0065] External Unstructured Data Acquisition: Utilizing distributed web crawling technology, unstructured text data streams related to mineral resources are captured in real time from multiple channels, including industry reports, commodity trading platforms, and policy announcements. This covers key information such as mineral categories, price fluctuations, supply and demand relationships, and policy guidance, comprehensively capturing market dynamics. Distributed web crawling technology, through the collaborative work of multiple crawler nodes, simultaneously crawls data from multiple target data sources, significantly improving data acquisition efficiency. A load balancing mechanism is employed to distribute crawling tasks, avoiding excessive load on individual nodes. Furthermore, anti-crawling mechanisms circumvent access restrictions from data sources, ensuring continuous and stable acquisition of external unstructured text data.

[0066] Unstructured data processing unit:

[0067] Data preprocessing: Denoising the captured unstructured text data, removing duplicate, invalid, and redundant information, and performing standardization operations to unify data format, encoding method, and language expression, thus clearing obstacles for subsequent processing;

[0068] Entity Relationship Extraction: A pre-trained natural language processing (NLP) model is used to perform deep analysis on the pre-processed text, accurately identifying key entities such as mineral resource categories, price fluctuation trends, supply and demand relationships, and policy guidance. Simultaneously, semantic relationships between entities are extracted, clarifying the intrinsic connections between various information elements. Based on the semantic understanding capabilities of the pre-trained model, the NLP model performs a series of operations, including word segmentation, part-of-speech tagging, entity recognition, and relation extraction, to mine key information and semantic relationships within the text. The pre-trained model, trained on large-scale text data, possesses powerful language understanding and information extraction capabilities, enabling it to accurately identify professional terms and complex semantic relationships in the field of mineral resources.

[0069] Dynamic knowledge graph construction: Based on the extracted entities and semantic relationships, a dynamically updated knowledge graph is constructed, connecting scattered information nodes through semantic relationships to form a structured knowledge network, which intuitively presents the correlation logic of various elements in market dynamics;

[0070] Market intelligence vector generation: Utilizing graph embedding algorithms, semantic information in knowledge graphs is mapped into low-dimensional, dense market intelligence vectors, transforming abstract textual knowledge into a computer-processable numerical form, thus achieving a structured representation of unstructured data. Graph embedding technology maps nodes and relationships in knowledge graphs to a low-dimensional vector space, transforming complex graph-structured data into dense numerical vectors while preserving the original semantic structure of the knowledge graph. By calculating the similarity between vectors, the degree of semantic association between entities can be quickly measured, providing an efficient numerical representation for subsequent data fusion and model processing.

[0071] Multi-source data fusion unit:

[0072] Multimodal data alignment: The generated market intelligence vector is aligned with structured data such as geological exploration data and mining operation data collected by sensor networks in a multimodal manner to establish the correspondence between different types of data and eliminate data heterogeneity;

[0073] Temporal correction and feature compensation: Utilizing a cross-modal attention mechanism, this system identifies temporal differences in data from different sources, corrects data with inconsistent timestamps, and ensures data synchronization over time. Simultaneously, it performs feature compensation to fill data gaps and ensure data integrity in cases of missing or anomalies. The cross-modal attention mechanism automatically focuses on information crucial to the fusion result in different modal data, highlighting the impact of key data by calculating the weight distribution of different modal features. During temporal correction, this mechanism identifies key time points of inconsistent data for accurate correction. Furthermore, during feature compensation, it focuses on relevant features of missing data to improve compensation accuracy.

[0074] Feature-level fusion coding: Through the feature-level fusion module, the corrected multi-source data is integrated and coded to generate a spatiotemporal tensor sequence that can be directly received by the unified digital twin model; each spatiotemporal unit contains fusion feature representations of geological attributes, equipment status, environmental indicators and market influencing factors, realizing a high degree of data integration.

[0075] In this embodiment, the unified digital twin model is the core hub of the dynamic prediction and mining management system for mineral resources. It receives multimodal data output from the data perception and fusion layer. Through deep fusion and dynamic processing of spatiotemporal features, it constructs a digital mirror that is highly synchronized with the real mining area, generating a comprehensive mining area situation map containing multi-dimensional information such as resources, risks, and economy. This provides accurate and real-time global status support for the dynamic prediction and decision-making intelligent agent, realizing the collaborative linkage between the real mining area and the digital space.

[0076] The core responsibility of the unified digital twin model is to receive, integrate, fuse features, and generate a situational awareness from multimodal data. It first standardizes the 3D voxel data of geological bodies, the time-series data of mining equipment operation status, environmental indicator monitoring point cloud data, and market intelligence vectors transmitted from the data perception and fusion layer, eliminating data discrepancies through spatiotemporal registration and multi-resolution fusion. Then, it extracts the fused spatiotemporal features using a 3D convolutional neural network with an embedded attention mechanism, ultimately decoding and generating a dynamically updated comprehensive mining area situational map. This situational map covers the probability distribution of resource reserves, grade gradient changes, geological risk factors, and economic indicators, mapping the real-time state of the mining area and providing comprehensive digital evidence for subsequent predictions and decisions. The unified digital twin model includes:

[0077] Multimodal data receiving and spatiotemporal registration unit:

[0078] Multimodal data reception: Real-time reception of various types of data from the data perception and fusion layer, including three-dimensional voxel data of geological bodies reflecting geological structure, time series of mining equipment operation status recording equipment operation status, environmental indicator monitoring point cloud data monitoring environmental changes, and market intelligence vectors reflecting market dynamics, ensuring the integrity and timeliness of data transmission;

[0079] Spatiotemporal registration processing: This process maps data from different sources onto a unified three-dimensional spatial coordinate system and time axis for the mining area. The three-dimensional spatial coordinate system is established based on the actual terrain and exploration data of the mine, clearly defining the physical location of each data point. The time axis uses the system's unified timestamp as a benchmark to synchronize the time dimension of all data, eliminating spatiotemporal deviations caused by different devices and interfaces, thus laying a unified spatiotemporal foundation for subsequent fusion. Spatiotemporal registration is based on the actual geographic information of the mine, establishing a globally unified three-dimensional spatial coordinate system. Through coordinate transformation algorithms, spatial data collected by different sensors and data interfaces are mapped to this coordinate system, eliminating spatial deviations caused by differences in equipment installation location and data acquisition perspective. In the time dimension, the timestamps of each data point are calibrated using the system's high-precision clock as a benchmark. Through time interpolation or delay compensation, the time records of data from different sources are made consistent with the actual event occurrence time, achieving unified alignment of data in the spatiotemporal dimension.

[0080] Multi-resolution fusion unit:

[0081] Data resolution analysis: Perform resolution detection on multimodal data after spatiotemporal registration to identify differences in sampling frequency (time dimension) and spatial granularity (spatial dimension) among different data; for example, three-dimensional voxel data of geological bodies may use a lower sampling frequency but a larger spatial granularity, time series of equipment operation status has a higher sampling frequency, and point cloud data of environmental indicator monitoring has uneven spatial granularity.

[0082] Spatiotemporal interpolation completion: A feature completion method based on spatiotemporal interpolation is adopted to align the sampling frequency and spatial granularity of various data. For the time dimension, low sampling frequency data is interpolated to make the data acquisition interval consistent with high sampling frequency data. For the spatial dimension, coarse-grained data is refined by interpolation, and uneven granularity data is uniformly adjusted. Finally, a multimodal spatiotemporal tensor with consistent sampling frequency and spatial granularity is generated to ensure the consistency of data in the spatiotemporal dimension.

[0083] Resolution fusion is based on spatiotemporal interpolation algorithms. Temporal interpolation uses linear interpolation or spline interpolation methods to calculate the estimated values ​​of missing time points based on known data from adjacent time points, enabling low-sampling-frequency data to reach the standard of high-sampling-frequency data. Spatial interpolation combines Kriging interpolation or inverse distance weighted interpolation, using known data from adjacent spatial points to refine coarse-grained or uneven-grained data and fill spatial gaps. Through interpolation operations, data of different resolutions are transformed into spatiotemporal tensors of uniform resolution, eliminating resolution barriers for subsequent feature fusion.

[0084] Spatiotemporal feature fusion unit:

[0085] Attention weight calculation: The multimodal spatiotemporal tensor is input into a three-dimensional convolutional neural network with an attention mechanism. The attention module dynamically analyzes the importance of different modal features in the spatiotemporal dimension and calculates the weight distribution of each feature. For geological body data related to geological anomaly areas, equipment operation data corresponding to changes in key equipment status, and market intelligence vectors related to large market fluctuations, higher weights are assigned to highlight the impact of core information on the mining area situation.

[0086] 3D Convolution Feature Extraction: The weighted multimodal spatiotemporal tensor is processed layer by layer through 3D convolution operations to extract deep spatiotemporal fusion features from the data. 3D convolution covers the length, width, and height dimensions of space as well as the time dimension, which can capture the changes in the spatial distribution of geological bodies over time, the temporal correlation of equipment status and the coupling of spatial location, the spatial diffusion of environmental indicators and the linkage of time, and the indirect spatiotemporal impact of market dynamics on various elements in the mining area, forming a comprehensive fusion feature representation.

[0087] The attention mechanism constructs a weight calculation model to analyze the correlation between each modal feature and key indicators of the mining area situation. The higher the correlation, the greater the feature weight, ensuring that core information dominates feature extraction. The three-dimensional convolutional neural network captures both the structural features of the spatial dimension and the changing features of the temporal dimension by sliding the three-dimensional convolutional kernel on the spatiotemporal data. The size and stride of the convolutional kernel are set according to the spatiotemporal scale of the mining area data, which can not only cover a sufficient spatiotemporal range to obtain global information, but also ensure the accuracy of feature extraction, ultimately achieving deep spatiotemporal fusion of multimodal data.

[0088] Integrated mining area situation map generation and dynamic update unit:

[0089] Multidimensional situation matrix decoding: The fused feature map output by the three-dimensional convolutional neural network is input into the fully connected layer and transformed into a multidimensional situation matrix containing multidimensional information through decoding operation; the matrix covers key contents such as the probability distribution of resource reserves (reflecting the possible range of resource quantity in different regions), grade gradient changes (reflecting the spatial distribution differences and trends of resource grade), geological risk factors (quantifying the magnitude and distribution of geological risks such as faults and water inrush), and economic indicators (regional economic value assessment combined with market and cost);

[0090] Time-series sliding window update: The comprehensive mining area situation map is dynamically updated based on a time-series sliding window mechanism. The sliding window operates at fixed time intervals. Every unit of time, the latest collected and processed multimodal data is incorporated into the analysis, feature fusion and situation decoding are re-performed, and various indicators in the multidimensional situation matrix are updated to ensure that the comprehensive mining area situation map and the actual mining area status changes remain synchronized in real time, achieving real-time synchronous rendering of the mining area status. The time-series sliding window has a fixed time window length and update step size. The window length is determined according to the response requirements of the mining area status changes, ensuring that the data within the window can reflect the complete status of the mining area in the recent period. The update step size corresponds to the data processing and analysis cycle. After each step of data analysis is completed, the window slides forward one step, discarding historical data that exceeds the window range and incorporating the latest real-time data. Through this dynamic sliding mechanism, the comprehensive mining area situation map is always generated based on the latest data, reflecting the dynamic changes in the mining area status in real time.

[0091] The workflow of the unified digital twin model is as follows:

[0092] Initialization phase:

[0093] Start the unified digital twin model and load the preset configuration information such as the three-dimensional spatial coordinate system parameters of the mining area, the time axis reference parameters, the three-dimensional convolutional neural network structure and initial weights, and the length and step size of the time-series sliding window;

[0094] Complete the communication connection test with the data perception and fusion layer and the dynamic prediction and decision-making intelligent agent to ensure that the data receiving and output channels are unobstructed. At the same time, initialize the data buffer and situation map storage unit to prepare for subsequent data processing.

[0095] Multimodal data reception and spatiotemporal registration phase:

[0096] Receive geological body three-dimensional voxel data, mining equipment operation status time sequence, environmental indicator monitoring point cloud data and market intelligence vector transmitted by the data perception and fusion layer, and temporarily store the data in the data cache area;

[0097] Perform a spatiotemporal registration operation on the data in the buffer area to map all data to a unified three-dimensional spatial coordinate system and time axis of the mining area, eliminate spatiotemporal deviations, and generate spatiotemporally aligned multimodal raw data;

[0098] Multi-resolution fusion stage:

[0099] Analyze the sampling frequency and spatial granularity of spatiotemporally aligned multimodal data, identify the resolution differences between different data, and determine the objects and ranges that need interpolation processing;

[0100] Spatiotemporal interpolation algorithms are used to complete and adjust low-resolution data, align the sampling frequency and spatial granularity of all data, generate a consistent multimodal spatiotemporal tensor, and store it in the feature processing unit.

[0101] Spatiotemporal feature fusion stage:

[0102] The multimodal spatiotemporal tensor is input into a three-dimensional convolutional neural network with an attention mechanism. The attention module calculates the spatiotemporal weight distribution of each modality feature, highlighting the features corresponding to the core influencing factors.

[0103] By extracting fusion features layer by layer through 3D convolution operations, the shallow single-modal features are gradually integrated into deep multimodal spatiotemporal fusion features, and the fusion feature map is output.

[0104] The generation and updating phase of the comprehensive mining area situation map:

[0105] The fused feature map is input into the fully connected layer and decoded to generate a multi-dimensional situation matrix containing the probability distribution of resource reserves, grade gradient changes, geological risk factors and economic indicators. Based on this matrix, an initial comprehensive mining area situation map is constructed.

[0106] According to the time-series sliding window mechanism, the latest multimodal data is incorporated every time the set update step size is reached. The spatiotemporal registration, multi-resolution fusion, and spatiotemporal feature fusion processes are repeated to re-decode and generate a multidimensional situation matrix, update the comprehensive mining area situation map, and achieve real-time synchronization.

[0107] Data output stage:

[0108] The dynamically updated comprehensive mining area situation map is transmitted in real time to the dynamic prediction and decision-making intelligent agent, providing a spatiotemporal sequence data source for the reserve-risk prediction sub-intelligent agent and a reference for the overall mining area status for the mining-economic optimization sub-intelligent agent;

[0109] At the same time, the updated comprehensive mining area status map is backed up and stored in the system database to record the historical changes in the mining area status for subsequent query, analysis and tracing.

[0110] End phase:

[0111] When the system receives a stop command or the mining operation is terminated, the unified digital twin model stops receiving new data, completes the data processing and situation map update within the current cycle, archives and stores the final comprehensive mining area situation map and historical data, closes communication connections with other modules, and releases the computing and storage resources occupied by the model operation.

[0112] In this embodiment, the dynamic prediction and decision-making intelligent agent is the "decision core" of the dynamic prediction and mining management system for mineral resources. Based on the comprehensive mining area situation map generated by the unified digital twin model, it realizes the dynamic prediction of resource reserves and geological risks, and the intelligent optimization of mining plans and economic strategies through the collaborative operation of the twin intelligent agents, providing accurate and efficient decision support for mining and balancing mining economy and safety.

[0113] The core responsibility of the dynamic prediction and decision-making agent is to complete two main tasks based on multi-dimensional information from the comprehensive mining area situation map: First, through the reserve-risk prediction sub-agent, it learns the patterns of spatiotemporal sequence data and dynamically outputs the probability distribution of resource reserves and geological risk levels of different zones within a specific future period; second, through the mining-economic optimization sub-agent, it constructs a dual-objective optimization model with the prediction results as constraints, and solves to generate the optimal mining path planning equipment scheduling scheme and resource allocation strategy. Simultaneously, the agent can respond to real-time data changes and user parameter adjustments, dynamically updating the decision results to ensure the adaptability, economy, and safety of the mining process.

[0114] The dynamic prediction and decision-making intelligent agent includes a reserve-risk prediction sub-agent and a mining-economic optimization sub-agent;

[0115] The reserve-risk prediction sub-agent is the core data output unit of the dynamic prediction and decision-making agent. Its main function is to extract key spatiotemporal data from the comprehensive mining area situation map generated by the unified digital twin model, learn the temporal change law of the mining area status through the gated cyclic unit network, and finally dynamically output the resource reserve probability distribution and geological risk level of different zones in a specific future period. It provides accurate and real-time core input data for the mining-economic optimization sub-agent and is a key bridge connecting the digital image of the mining area and decision optimization.

[0116] The core tasks of the reserve-risk prediction sub-agent revolve around four stages: data processing, model learning, prediction output, and dynamic updating. First, it filters and processes multi-dimensional spatiotemporal data related to resources and risks from the comprehensive mining area situation map, eliminating data discrepancies and interference. Then, it mines the temporal dependencies in the data through a gated recurrent unit network, learning the inherent laws governing changes in resource reserves and the evolution of geological risks. Subsequently, it transforms the learning results into a quantitative probability distribution of reserves and a qualitative risk level. Finally, it continuously receives updated data from the comprehensive mining area situation map, repeating the above process to achieve real-time iteration of the prediction results, ensuring that the output data always remains synchronized with the actual state of the mining area, providing a reliable basis for subsequent mining scheme optimization. The reserve-risk prediction sub-agent includes:

[0117] Multidimensional spatiotemporal data extraction and preprocessing:

[0118] Data extraction: Multidimensional spatiotemporal sequence data are screened and extracted from the comprehensive mining area situation map generated by the unified digital twin model. These data cover the temporal changes of resource reserves, the spatiotemporal fluctuations of grade distribution, the dynamic evolution of geological risk factors, and the time correlation characteristics of economic indicators, comprehensively covering the key elements affecting resource reserve prediction and geological risk assessment.

[0119] Temporal alignment: Temporal alignment is performed on the extracted multidimensional spatiotemporal sequence data. Since the collection frequency and recording time nodes of different data may vary, temporal alignment is necessary to ensure that the time dimension of all data is consistent, so that each set of data can accurately correspond to the mining area status of the same time segment and avoid the interference of time deviation on subsequent model learning.

[0120] Normalization: Normalize the time-series aligned data. Different types of data have significantly different units. For example, the unit of resource reserves may be 10,000 tons, and the geological risk factor may be a dimensionless exponential value. Directly inputting these into the model will cause the impact of some data to be amplified or reduced. Through normalization, all data are mapped to a unified numerical range, eliminating the difference in units, ensuring data consistency, and providing standardized learning samples for the model.

[0121] Temporal dependency learning based on gated recurrent unit networks:

[0122] Model input: The standardized spatiotemporal sequence data is input into the gated recurrent unit network; this network is a deep learning model specifically designed for long time series data, which can effectively handle the correlation of data in the time dimension, and is especially suitable for scenarios such as mining areas where the state changes slowly over time and there is long-term dependence.

[0123] Gating mechanism operation: The gated recurrent unit network achieves time-dependent learning through two core structures: the reset gate and the update gate. The reset gate determines whether to ignore invalid information in historical data. When the historical data of a certain period has little impact on the prediction of the current state, the reset gate will weaken the weight of that part of the data. The update gate controls the degree of contribution of historical data to the current state. For features with continuous characteristics such as the trend of resource reserve changes and the accumulation of geological risks, the update gate will retain more of the influence of historical data to ensure that the model can capture long-term change patterns.

[0124] Hidden state output: Through the synergistic effect of the reset gate and the update gate, the gated recurrent unit network learns the standardized spatiotemporal sequence data layer by layer, and finally outputs a set of hidden state sequences. This set of sequences contains the key feature information of the data in the time dimension, covering both the short-term fluctuation details of resource reserves and geological risks, and preserving the long-term trend of change. It is the core basis for generating subsequent prediction results.

[0125] The core advantage of gated recurrent unit networks lies in solving the gradient vanishing or exploding problems that traditional recurrent neural networks encounter when processing long-term data through a gating mechanism. Both the reset gate and the update gate use the sigmoid activation function to output values ​​between 0 and 1, where 0 indicates that the corresponding information is completely ignored and 1 indicates that the corresponding information is completely retained. When processing mining area time-series data, the reset gate filters out abnormal data that occasionally occurs during the mining process (such as false reserve changes caused by a temporary equipment failure), while the update gate focuses on retaining long-term trend information such as gradual changes in resource grade and accumulation of geological risks. Through this selective learning, the network can accurately capture the effective temporal dependencies in the data, ensuring the accuracy of predictions for resource reserves and geological risks over long periods.

[0126] Resource reserve probability distribution and geological risk level generation:

[0127] Resource reserve probability distribution generation: Based on the hidden state sequence output by the gated recurrent unit network, the hidden state is first transformed through a fully connected layer, mapping high-dimensional features to low-dimensional features related to reserve prediction. Then, the transformed features are input into the probability output layer, which uses the softmax function to calculate the probability of different reserve intervals, and finally generates the resource reserve probability distribution of different partitions in a specific future period. This distribution can intuitively present the probability of the resource quantity of each partition being in different intervals. For example, the probability of a certain partition having resource reserves of 1 million to 1.2 million tons is 60%, and the probability of having reserves of 800,000 to 1 million tons is 30%, providing a quantitative resource basis for subsequent optimization decisions.

[0128] Geological risk level generation: Simultaneously, based on the hidden state sequence, it is input into the risk classification layer; the risk classification layer transforms the geological risk-related features in the hidden state into predefined risk categories through preset classification rules and feature mapping relationships; the predefined risk categories typically include three levels: low risk, medium risk, and high risk. The classification process comprehensively considers the change range and duration of geological risk factors in the hidden state, as well as the similarity with historical high-risk states, and finally outputs the geological risk level corresponding to each partition, providing a clear risk basis for the subsequent construction of safety constraints;

[0129] Prediction results are updated dynamically:

[0130] Real-time data reception: The reserve-risk prediction sub-intelligent body continuously monitors the update signals of the unified digital twin model. Once new data is generated in the integrated mining area situation map (such as new geological exploration results or real-time mining operation data updates), it immediately receives the new data to ensure that the learning samples are always based on the latest mining area status.

[0131] Iterative execution of the process: After receiving new data, the sub-agent repeats the complete process of "data extraction and preprocessing - temporal dependency learning - prediction result generation"; the new data will be integrated into the standardized spatiotemporal sequence, the gated recurrent unit network will be retrained, the hidden state sequence will be updated, and then a new resource reserve probability distribution and geological risk level will be generated.

[0132] Real-time output of results: The updated prediction results will be transmitted to the mining-economic optimization sub-agent as soon as possible, and stored in the system database for traceability. This dynamic update mechanism can ensure that the prediction results keep pace with the changes in the actual state of the mining area. For example, when the reserves of a certain area decrease due to mining or the risk increases due to geological changes, the prediction results will quickly reflect these changes and provide real-time support for decision optimization.

[0133] The reserve-risk prediction sub-agent adopts a dual-output mode of "quantitative + qualitative". It quantitatively describes the uncertainty of resource reserves through probability distribution and qualitatively classifies the severity of geological risks through risk levels. The core principle of this mode is to perform multi-dimensional feature mapping on the hidden state output by the gated recurrent unit network: for resource reserves, the features are mapped to a probability distribution through a fully connected layer and a softmax function, highlighting the numerical uncertainty of reserves; for geological risks, the features are mapped to discrete risk levels through a risk classification layer, highlighting the decision-making nature of risks. Both are generated based on the same hidden state sequence, ensuring the inherent consistency between resource and risk prediction results and avoiding logically contradictory prediction results such as "high reserves and extremely high risks".

[0134] The mining-economic optimization sub-agent is the core decision-making output unit of the dynamic prediction and decision-making agent. Its main function is to receive the resource reserve probability distribution and geological risk level output by the reserve-risk prediction sub-agent. With the goal of maximizing net present value and minimizing mining safety accident risk, it constructs a mixed integer programming model. Through multi-objective optimization, it generates the optimal mining path plan, equipment scheduling scheme and resource allocation strategy. At the same time, it can dynamically adjust the decision scheme according to the real-time status changes of the mining area, realize the synergistic optimization of the economy and safety of the mining process, and provide directly implementable operational guidance for actual mining operations in the mining area.

[0135] The core tasks of the mining-economic optimization sub-agent revolve around five stages: input processing, model building, optimization solution, scheme output, and dynamic adjustment. First, the output data of the reserve-risk prediction sub-agent is transformed into constraints that the model can recognize. Then, based on the actual operational needs of the mining area, a dual-objective optimization model balancing economic benefits and safety risks is built. Subsequently, a multi-objective optimization algorithm and solver are used to solve the model, selecting the optimal solution that meets the actual needs. Next, detailed mining, scheduling, and allocation schemes are generated based on the optimal solution. Finally, the model parameters and decision-making schemes are continuously updated with real-time data and prediction results from the mining area, ensuring that the output scheme always adapts to the current state of the mining area, avoiding resource waste and economic losses while preventing safety accidents. The mining-economic optimization sub-agent includes:

[0136] Input data processing and constraint construction:

[0137] Resource reserve constraint construction: Receive the resource reserve probability distribution output by the reserve-risk prediction sub-agent, and use probabilistic statistical methods to calculate the quantile estimate of the resource reserve. The quantile estimate can reflect the reserve boundary at different confidence levels. For example, the 90th quantile is used as the upper limit constraint of the reserve and the 10th quantile is used as the lower limit constraint of the reserve. This defines the reasonable range of exploitable resources in each zone, avoiding over-exploitation (leading to the risk of resource depletion) or under-exploitation (causing economic loss) caused by over-reliance on a single reserve prediction value. It also fully considers the uncertainty of resource reserves and sets a scientific resource boundary for model optimization.

[0138] Safety risk constraint construction: The system receives the geological risk level output by the reserve-risk prediction sub-agent and transforms the qualitative risk level into a quantitative risk probability value through a risk mapping function. For example, low risk level corresponds to a safety accident probability of less than 5%, medium risk level corresponds to a safety accident probability of 5%-15%, and high risk level corresponds to a safety accident probability of more than 15%. These risk probability values ​​are used as safety constraints for the model. Simultaneously, historical equipment failure data (such as equipment failure rate over the past year), human operation error coefficients (such as statistical values ​​of operation errors from different shifts), and environmental monitoring indicators (such as the frequency of exceeding standards for underground humidity and gas concentration) are incorporated to further quantify potential risks during the mining process. This ensures that the constraints comprehensively cover all types of safety influencing factors and avoids constraint loopholes caused by a single risk indicator.

[0139] Construction of dual objective function:

[0140] Net Present Value (NPV) Maximization Objective Function Construction: NPV is a core indicator for measuring the long-term economic benefits of mining projects. Its calculation dimensions include resource extraction volume, market price forecasts, discount rates, equipment operating costs, and environmental compliance costs. Resource extraction volume is based on the constraints of recoverable reserves in each zone, combined with mining efficiency to set a reasonable range. Market price forecasts integrate dynamic market data obtained from data sensing and fusion layers (such as the price fluctuation trend of mineral products in the past three months and the impact of policy guidance on prices in the next six months) to ensure that price parameters are synchronized with market changes. The discount rate is determined with reference to the industry average cost of capital and the project risk coefficient, reflecting the time value of money. Equipment operating costs include equipment depreciation, energy consumption, and maintenance costs, dynamically calculated based on equipment model and operating time. Environmental compliance costs cover mine reclamation, pollutant treatment, and other expenses, complying with environmental protection policy requirements. By integrating these parameters into the NPV calculation logic, a NPV maximization objective function is formed, ensuring that the model optimization direction is consistent with the long-term economic benefit goals of the mining area.

[0141] Construction of the objective function for minimizing mining safety accident risk: Risk calculation is based on the risk probability value in the safety risk constraints, while also integrating historical equipment failure data, human operation error coefficients, and environmental monitoring indicators. For example, the weight of geological risk probability value is set to 40%, equipment failure rate to 25%, human operation error coefficient to 20%, and environmental indicator exceedance frequency to 15%. The comprehensive safety risk value of each zone is calculated by weighted summation. Minimizing the comprehensive safety risk value is used as the objective function, which highlights the impact of geological risk as the core safety factor, while also taking into account secondary safety factors such as equipment, personnel, and environment. This ensures that model optimization can comprehensively reduce the probability of safety accidents during the mining process and avoid accident hazards caused by neglecting certain safety factors.

[0142] Mixed-integer programming models can handle both discrete and continuous decision variables simultaneously, adapting to the complex decision-making needs of mining operations. Discrete decision variables correspond to non-continuous decisions such as mining path selection (e.g., the order of mining zones) and equipment scheduling (e.g., whether equipment is allocated to a certain area). Continuous decision variables correspond to continuous decisions such as resource extraction (e.g., how many tons are extracted per month in a certain zone) and resource allocation (e.g., how much funding is allocated to a certain project). By setting integer constraints (e.g., equipment numbers and zone numbers are integers) and continuous constraints (e.g., extraction volume and funding are continuous values), the model integrates various decision requirements into a unified mathematical model, ensuring that the solution results can satisfy both the determinism of discrete decisions and the refinement of continuous decisions, avoiding the problem that traditional single-type variable models cannot cover complex decision scenarios.

[0143] Multi-objective optimization solution:

[0144] The dual-objective problem is transformed into a single-objective problem by using a multi-objective optimization algorithm to maximize net present value (NPV) and minimize mining safety risk. A Pareto front generation method is employed to generate a set of non-dominated solutions, satisfying all constraints. A non-dominated solution is one that cannot improve one objective without decreasing the other; for example, a solution that increases NPV by 5% without increasing safety risk is a non-dominated solution. Subsequently, the non-dominated solutions are scored and ranked based on the preference weights of the mine's decision-makers (e.g., 60% for NPV and 40% for safety in economy-priority mines, and vice versa for safety-priority mines), and the selected solutions are then selected. The optimal candidate solution with the highest comprehensive score; the core of the Pareto front generation method is to find the "optimal equilibrium solution set" in bi-objective optimization; since there is a certain conflict between net present value and safety risk; increasing mining volume may increase net present value, but may also increase safety risk, the Pareto front can find all "optimal equilibrium points", i.e., non-dominated solutions, in the conflict; the non-dominated solutions generated by the Pareto front cover different economic-safety equilibrium states, and the optimal solution is selected by combining the decision-maker's preference weights, which avoids the extreme cases of "emphasizing economy over safety" or "emphasizing safety over economy" caused by single objective optimization, and can fully fit the actual operation orientation of the mining area, ensuring the practicality and flexibility of the optimization results;

[0145] Feasibility Verification and Local Optimization: The optimal candidate solution is input into the mixed integer programming solver to verify its feasibility. The solver checks whether the candidate solution meets all constraints, such as whether the extraction volume is within the reserve constraint range, whether the comprehensive safety risk value is lower than the set threshold, and whether there are time conflicts in equipment scheduling. If there are cases where the constraints are not met, the solver will perform local optimization adjustments on the candidate solution, such as reducing the extraction volume in high-risk areas, adjusting equipment operation periods to avoid conflicts, and reallocating resources to balance the needs of each link, until the candidate solution fully meets all constraints, and outputs the final optimization result.

[0146] The constraint construction process incorporates the uncertainty of prediction results and the dynamism of real-time data; quantile estimation is used to process the probability distribution of resource reserves, enabling reserve constraints to reflect resource boundaries at different confidence levels and adapt to the uncertainty of resource prediction; geological risk levels are transformed through risk mapping functions and multiple safety indicators are integrated, ensuring that safety constraints comprehensively cover dynamically changing safety factors; simultaneously, when the real-time status of the mining area changes, the constraints are adjusted synchronously with the prediction results and real-time data input. For example, when the actual mining volume of a certain zone decreases, the upper limit of the reserve constraint for that zone will be lowered to ensure that the constraints are always consistent with the current status of the mining area and to avoid model optimization deviating from reality due to the solidification of constraints;

[0147] Decision-making scheme generation and dynamic adjustment:

[0148] Decision-making scheme generation: Based on the final optimization results, three types of detailed schemes are generated; the mining path planning sequence clarifies the mining order, mining scope, and mining progress of each zone, for example, mining zone A with sufficient reserves and low risk first, and then mining zone B with medium reserves and medium risk, while marking the mining depth and working face layout of each zone; the equipment scheduling schedule allocates the working time, working area, and maintenance cycle of each piece of equipment according to the mining progress and equipment performance, for example, allocating large mining equipment to zones with larger mining volumes and setting the equipment maintenance period from 2:00 AM to 4:00 AM every day; the resource allocation strategy optimizes the allocation of human, material, and financial resources, for example, equipping high-risk zones with more safety monitoring personnel, increasing spare parts reserves in areas with dense equipment, and prioritizing the allocation of funds for key mining projects;

[0149] Dynamic adjustment: Continuously receive update signals from two aspects: first, real-time status data of the mining area transmitted by the unified digital twin model (such as the actual mining volume of a certain zone exceeding the predicted value, or sudden equipment failure); second, the prediction results updated by the reserve-risk prediction sub-agent (such as the geological risk level of a certain zone increasing). When an update signal is received, the sub-agent readjusts the constraints (such as narrowing the reserve constraint range of the risk-increased zone), corrects the objective function parameters (such as increasing the weight of the operating cost of the faulty equipment), repeats the multi-objective optimization solution process, generates a new decision scheme, and replaces the old scheme in real time to ensure that the scheme is always adapted to the current state of the mining area and avoids economic losses or safety risks caused by the lag in the scheme.

[0150] In this embodiment, the visualization and interactive control layer is the core of human-computer interaction in the dynamic prediction and mining management system for mineral resources. It connects the dynamic prediction and decision-making intelligent agent, and uses three-dimensional visualization technology to intuitively present the digital twin status of the mining area and the optimal decision-making scheme. At the same time, it provides risk warning and parameter adjustment interfaces to help users quickly understand the mining area situation, accurately adjust decision-making strategies, and achieve efficient mineral management through human-computer collaboration.

[0151] The core responsibility of the visualization and interactive control layer is to receive full-dimensional data output by the dynamic prediction and decision-making intelligent agent, including comprehensive mining area situation maps, resource reserve probability distribution, geological risk levels, optimal mining path planning, and equipment scheduling schemes. Through data parsing and format conversion, this data is transformed into structured scene data adapted to the 3D rendering engine. Then, a dynamic 3D mining area scene is constructed through real-time graphics rendering, and a risk warning engine is integrated to generate visualized warning signals. Simultaneously, an interactive control panel is deployed, allowing users to adjust decision parameters and trigger system re-optimization. Finally, a multi-view collaborative mode assists users in comprehensive decision-making, ensuring that the visualized content is synchronized with the latest decision data in real time, improving decision-making efficiency and accuracy. The visualization and interactive control layer includes:

[0152] Data parsing and format conversion unit:

[0153] Data reception: Real-time reception of various output data transmitted by the dynamic prediction and decision-making intelligent agent, including multi-dimensional situation matrix of comprehensive mining area situation map, interval probability data of resource reserve probability distribution, zonal risk classification results of geological risk level, sequence coordinate data of optimal mining path planning and time and location correlation information of equipment scheduling scheme, to ensure the integrity and timeliness of data reception.

[0154] Data analysis: The received data is analyzed in layers to extract key information such as the spatial coordinates of geological bodies, resource grade values, risk level indicators, mining path nodes, and equipment operating parameters. The correlation logic between the data is sorted out, for example, the mining path nodes are matched and associated with the resource reserves and risk levels of the corresponding zones.

[0155] Format conversion: Convert the parsed unstructured or semi-structured data into structured scene data formats that can be recognized by the 3D rendering engine, including voxelized data formats of geological bodies, wireframe model data formats of mining paths, animation model data formats of equipment, and point cloud data formats of environmental indicators, to ensure that the data can be directly called by the rendering engine.

[0156] Dynamic 3D scene rendering unit:

[0157] Rendering pipeline construction: Based on structured scene data, the real-time graphics rendering pipeline is started to complete basic rendering configurations such as scene initialization, model loading, lighting settings, and material attachment, and to build a 3D scene framework that conforms to the actual geographical features of the mining area, ensuring that the spatial proportion of the scene is consistent with the real mining area.

[0158] Geological body visualization: Voxelization rendering technology is used to process geological body data and color gradient is used to map the distribution of resource grades. For example, dark colors are used for high-grade resource areas and light colors are used for low-grade resource areas. The differences in resource quality in different areas are presented intuitively through the color depth. At the same time, users can view the internal structure of geological bodies by zooming and rotating.

[0159] Visualization of mining paths and equipment: The optimal mining path planning data is transformed into a dynamically highlighted path line, which is overlaid and displayed at the corresponding position in the 3D scene. The path line color can distinguish mining priority. Load the equipment animation model and dynamically simulate the working status of the equipment in different areas according to the time nodes of the equipment scheduling plan, such as digging and transportation, to intuitively display the equipment scheduling process.

[0160] Environmental indicator visualization: By fusing environmental monitoring data with point cloud data, a heat map of environmental indicators is generated. For example, the air quality indicator is represented by a red, yellow and green heat map to indicate the degree of pollution, and the soil moisture indicator is represented by a blue layer with different transparency to indicate the degree of moisture. These are superimposed on the corresponding spatial areas of the 3D scene to reflect changes in environmental status in real time.

[0161] The 3D scene rendering is based on a real-time graphics rendering pipeline. It transforms the 3D spatial data of geological bodies into discrete voxel units through voxel rendering. Each voxel unit carries attribute information such as resource grade and risk level. Then, through color mapping and lighting calculation, it presents the spatial morphology and attribute differences of geological bodies. For mining paths and equipment, a combination of wireframe models and animation models is used. Through coordinate transformation and frame animation control, the path is dynamically displayed and the equipment action is simulated, ensuring the realism and interactivity of the 3D scene.

[0162] Risk warning engine unit:

[0163] Risk data extraction: Extract the zoning and classification results of geological risk levels and the probability calculation data of mining safety accident risks from the parsed dynamic prediction and decision-making intelligent agent data, and clarify the basic risk information of each spatial unit;

[0164] Comprehensive risk value calculation: Through risk mapping algorithm, geological risk and mining safety accident risk data are integrated, and combined with factors such as resource mining intensity and equipment distribution density of each spatial unit, the comprehensive risk value of each spatial unit is calculated to quantify the severity of risk;

[0165] Visualized early warning signals: Based on the comprehensive risk value, visualized early warning signals are generated using multiple methods; high-risk areas are highlighted with a red gradient, medium-risk areas are represented by a yellow gradient, and low-risk areas are marked with a green gradient; flashing indicators are added to high-risk areas to alert users through intermittent illumination; for areas with a wide risk range, risk isosurfaces are generated to intuitively present the risk spread range and gradient changes, ensuring that users can quickly locate high-risk areas;

[0166] Risk visualization is based on color coding and spatial identification theory, transforming abstract risk values ​​into intuitive visual signals. By establishing a mapping relationship between risk values ​​and colors, high risk corresponds to high-saturation warning colors, and low risk corresponds to low-saturation safety colors, using human sensitivity to color to quickly convey risk information. At the same time, it combines flashing signs and isosurface technology. Flashing signs enhance the attention to high-risk areas through visual dynamic changes, while isosurfaces present the spatial distribution gradient of risk through spatial curved surface shapes, realizing a multi-dimensional visual expression of risk information.

[0167] Interactive control panel unit:

[0168] Parameter Adjustment Interface Deployment: Design and deploy a multi-dimensional decision parameter adjustment interface, including a risk tolerance slider, an economic weight input box, and a mining priority selector; the risk tolerance slider allows users to adjust it within the range of 0 to 100, with higher values ​​indicating a higher tolerance for risk; the economic weight input box allows users to input values ​​from 0 to 1, with higher values ​​indicating a greater emphasis on economic benefits during optimization; the mining priority selector provides options such as resource grade priority, risk avoidance priority, and equipment efficiency priority, allowing users to choose according to their actual needs;

[0169] Parameter feedback and re-optimization: After the user modifies the parameters in real time through the graphical interface, the system automatically encapsulates the adjusted parameters into new constraints and transmits them to the dynamic prediction and decision-making agent; after receiving the new constraints, the agent restarts the optimization solution process, generates a decision scheme adapted to the new parameters, and realizes the closed-loop linkage between parameter adjustment and decision optimization.

[0170] Interactive control is based on a parameter feedback closed-loop mechanism. The parameters input by the user through the control panel are encapsulated as constraints and transmitted to the dynamic prediction and decision-making agent, triggering the agent's re-optimization process. After the agent outputs a new decision scheme, the incremental update unit only refreshes the changed data locally, avoiding a full scene redraw. This reduces system resource consumption and ensures real-time linkage between user operation and scene update, realizing a complete closed loop of "user operation - parameter transmission - optimization solution - scene update".

[0171] Incremental update unit:

[0172] Update signal monitoring: Continuously monitor the optimization result update signal of the dynamic prediction and decision-making agent. When the user adjusts the parameters to trigger re-optimization, or when the real-time data of the mining area changes and the agent outputs a new decision scheme, the update signal is captured in time and the incremental update process is started.

[0173] Local data refresh: There is no need to re-render the entire 3D scene. Only the changed data is updated locally. For example, after the mining path is adjusted, only the path line data is refreshed. After the risk level changes, only the color and flashing status of the corresponding area are updated. After the equipment scheduling plan is modified, only the time and position parameters of the equipment animation model are updated, which reduces the consumption of rendering resources and ensures the real-time update of the scene.

[0174] Data synchronization verification: After the update is completed, the scene data before and after the update is compared with the data output by the intelligent agent to verify the consistency between the visualization content and the latest decision data, and to ensure that the information such as mining path, equipment status, and risk warning in the 3D scene is completely matched with the latest decision results of the intelligent agent.

[0175] Multi-view collaborative visualization unit:

[0176] Multi-view deployment: Supports users to browse multiple views simultaneously, including a 3D mining area scene view, a 2D plan view, and a time-series risk curve view; the 3D scene view displays the three-dimensional state of the mining area, the 2D plan view presents the planar layout and data distribution of the mining area in a top-down view, and the time-series risk curve view displays the trend of risk values ​​in each zone over time with time as the horizontal axis and risk value as the vertical axis.

[0177] View Linkage and Data Drill-down: Through view linkage technology, interactive connections between different views are realized. For example, clicking on a certain partition in a 3D scene will simultaneously highlight that partition in the 2D planar view, and the time series risk curve view will automatically locate and enlarge the risk curve of that partition. At the same time, users can perform data drill-down in any view. Double-clicking on a certain area will allow users to view detailed data for that area, including specific values ​​of resource reserves, the basis for risk level calculation, detailed equipment operating parameters, etc., to help users conduct in-depth analysis and decision-making information.

[0178] Multi-view collaboration is based on data association and event-driven mechanisms. It establishes a unified data index for different views, with each data index corresponding to a spatial unit or time node in the mining area, ensuring that the data called by different views comes from the same data source. When a user triggers an interactive event in a view, such as clicking or double-clicking, the system updates the display status of the corresponding data index in other views synchronously through the event-driven mechanism, realizing view linkage. At the same time, through data drill-down technology, it retrieves detailed underlying data based on the data index, meeting the user's needs from macro-level browsing to micro-level analysis.

[0179] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic prediction and mining management system for mineral resources, characterized in that: include: The data perception and fusion layer collects structured geological exploration data, mining operation data, environmental monitoring data, and unstructured market dynamic data in real time through a sensor network deployed at the mine site and external data interfaces; it uses an entity relationship extraction model based on natural language processing to structure the unstructured market dynamic data and generate market intelligence vectors. A unified digital twin model for receiving and processing multimodal data; The unified digital twin model uses a three-dimensional convolutional neural network with an embedded attention mechanism to fuse the spatiotemporal features of geological bodies, mining equipment, environmental indicators, and market intelligence vectors, and outputs a dynamically updated comprehensive mining area status map that includes resource reserves, grade distribution, geological risk factors, and economic indicators. A dynamic prediction and decision-making intelligent agent is used for collaborative decision-making based on a comprehensive mining area situation map; the dynamic prediction and decision-making intelligent agent includes: Reserves-Risk Prediction Sub-Agent: Multidimensional spatiotemporal sequence data is extracted from the comprehensive mining area situation map generated by the unified digital twin model. The multidimensional spatiotemporal sequence data includes the temporal changes of resource reserves, grade distribution, geological risk factors and economic indicators. The extracted multidimensional spatiotemporal sequence data is preprocessed with temporal alignment and normalization to eliminate dimensional differences and ensure data consistency, thereby generating a standardized spatiotemporal sequence. The standardized spatiotemporal sequence is input into the gated recurrent unit network, and the temporal dependencies are learned through the gating mechanism, and the hidden state sequence is output. Based on the hidden state sequence, a probability distribution of resource reserves in different partitions within a specific future period is generated through a fully connected layer and a probability output layer. The probability output layer uses a softmax function to represent the probability of different reserve intervals. Meanwhile, based on the hidden state sequence, a geological risk level is generated through a risk classification layer, where the risk classification layer maps the hidden state to a predefined risk category; The probability distribution of resource reserves and geological risk levels are dynamically updated to reflect real-time data changes and provide input for the mining-economic optimization sub-agent; Mining-Economic Optimization Sub-Agent: Based on the probability distribution of resource reserves, quantile estimates of resource reserves are calculated using probabilistic statistical methods, and these quantile estimates are used as the resource reserve constraint boundary of the mixed integer programming model to reflect the impact of resource uncertainty. Based on geological risk levels, risk levels are transformed into corresponding risk probability values ​​through a risk mapping function, and the risk probability values ​​are used as safety risk constraints in a mixed-integer programming model to quantify potential risks in the mining process. A net present value (NPV) maximization objective function is constructed, where NPV calculation is based on resource extraction volume, market price forecasts, discount rates, equipment operating costs, and environmental compliance costs, and market dynamic data is integrated to enhance the real-time nature of economic assessment. An objective function for minimizing mining safety accident risks is constructed, in which risk calculation is based on geological risk probability values, historical equipment failure data, human operation error coefficients, and environmental monitoring indicators to comprehensively cover safety factors; The bi-objective function is transformed into a solvable single-objective optimization problem through a multi-objective optimization algorithm. The Pareto front generation method is used to generate a set of non-dominated solutions, and the optimal candidate solution is selected from the non-dominated solutions based on the decision-maker's preference weights. The optimal candidate solution is verified for feasibility and locally optimized using a mixed-integer programming solver. The mining sequence and equipment allocation are adjusted to meet all constraints, and the final optimization result is output. Based on the final optimization results, a detailed mining path planning sequence, equipment scheduling schedule and resource allocation strategy are generated, and the strategy is dynamically updated to respond to real-time data changes to ensure the adaptability and efficiency of the mining process. The visualization and interactive control layer connects to the dynamic prediction and decision-making intelligent agent. It is used to render the unified digital twin model and the optimal mining path planning in a three-dimensional visualization form, and to provide users with visualized early warning signals and decision parameter adjustment interfaces based on geological risk level and mining safety accident risk.

2. The mineral resource dynamic prediction and mining management system according to claim 1, characterized in that: The data perception and fusion layer includes the processing of unstructured market dynamic data and the fusion of multi-source data, which includes: Unstructured text data streams are obtained in real time from industry reports, commodity trading platforms and policy announcements through distributed web crawlers, and the obtained text data is preprocessed by denoising and standardization. A pre-trained natural language processing model is used to perform entity recognition and relation extraction on the pre-processed text, extracting key entities and their semantic relationships, including mineral resource categories, price fluctuation trends, supply and demand relationships, and policy orientations. A dynamic knowledge graph is constructed based on the extracted entity relationships, and the semantic information in the knowledge graph is mapped into a low-dimensional dense market intelligence vector through a graph embedding algorithm. The market intelligence vector is aligned with geological exploration data and mining operation data collected by sensor networks in a multimodal manner, and a cross-modal attention mechanism is used to correct timestamps and compensate features for data with inconsistent time series. The corrected multi-source data is encoded into a spatiotemporal tensor sequence that can be received by the unified digital twin model through the feature-level fusion module. Each spatiotemporal unit contains a fusion feature representation of geological attributes, equipment status, environmental indicators, and market influencing factors.

3. The mineral resource dynamic prediction and mining management system according to claim 1, characterized in that: The unified digital twin model receives and processes multimodal data to achieve spatiotemporal feature fusion and dynamic generation of a comprehensive mining area situation map, including: It receives multimodal data from the data perception and fusion layer, including three-dimensional voxel data of geological bodies, time series of mining equipment operation status, point cloud data of environmental indicator monitoring, and market intelligence vectors. It performs preliminary spatiotemporal registration on the multimodal data and maps data from different sources onto a unified three-dimensional spatial coordinate system and time axis of the mining area. Multi-resolution fusion is performed on the spatiotemporally registered multimodal data. The sampling frequency and spatial granularity of geological bodies, equipment status, environmental indicators and market intelligence vectors are aligned by a feature completion method based on spatiotemporal interpolation to generate a consistent multimodal spatiotemporal tensor. A three-dimensional convolutional neural network with an attention mechanism is embedded with a multimodal spatiotemporal tensor input. The attention module dynamically calculates the weight distribution of different modal features in the spatiotemporal dimension to highlight the impact of geological anomalies, changes in key equipment status, and market fluctuations. The fused spatiotemporal feature representation is extracted layer by layer through three-dimensional convolution operations. The output feature map of a three-dimensional convolutional neural network is decoded into a multi-dimensional situation matrix containing the probability distribution of resource reserves, grade gradient changes, geological risk factors and economic indicators through a fully connected layer. The comprehensive mining area situation map is dynamically updated based on a time-series sliding window mechanism to achieve real-time synchronous rendering of the mining area status.

4. The mineral resource dynamic prediction and mining management system according to claim 1, characterized in that: The visualization and interactive control layer is used to render the unified digital twin model and optimal mining path planning in a 3D visualization format, and provides users with visualized early warning signals and decision parameter adjustment interfaces based on geological risk levels and mining safety accident risks, including: It receives output data from the dynamic prediction and decision-making intelligent agent, including comprehensive mining area situation map, resource reserve probability distribution, geological risk level, optimal mining path planning and equipment scheduling scheme, and parses and converts the output data to generate structured scene data suitable for 3D rendering engine. Based on structured scene data, a dynamic 3D mining scene is constructed through a real-time graphics rendering pipeline. The geological body uses voxel rendering technology to represent the distribution of resource grades with color gradients. The mining path planning is displayed by overlaying dynamic highlighted path lines and equipment animation models. Environmental indicators are generated by fusion of point cloud data to form a thermal layer. An integrated risk warning engine calculates the comprehensive risk value of each spatial unit based on the parsed geological risk level and mining safety accident risk data through a risk mapping algorithm. The warning signal is then rendered in a 3D scene using color coding, flashing markers, and risk isosurfaces, with high-risk areas highlighted by a red gradient. Deploy an interactive control panel to provide users with a multi-dimensional decision parameter adjustment interface, including a risk tolerance slider, an economic weight input box, and a mining priority selector. After users modify the parameters in real time through the graphical interface, the system automatically encapsulates the adjusted parameters as constraints and feeds them back to the dynamic prediction and decision-making agent for re-optimization. Based on the re-optimization results triggered by user parameter adjustments, the mining path planning, equipment scheduling scheme and early warning signals in the 3D scene are dynamically refreshed through an incremental update mechanism to ensure that the visualization content is synchronized with the latest decision data; It supports a multi-view collaborative visualization mode, allowing users to simultaneously browse 3D scenes, 2D planar plots, and time-series risk curves. Through view linkage technology, it enables data drill-down and interactive exploration from any perspective to assist users in comprehensive decision analysis.

Images