Surface mine geological parameter optimization modeling method based on deep reinforcement learning

Abstract

The invention discloses a surface mine geological parameter optimization modeling method based on deep reinforcement learning, and the method comprises the steps: achieving the spatial unified expression of multi-source data through constructing a three-dimensional non-uniform grid frame, and obtaining the spatial attribute description with a hierarchical discrimination degree on the basis of embedding a multi-scale feature extractor; and the consistency of initial parameter distribution and historical observation data is improved in combination with a dynamic weight distribution mechanism. According to the method, a strategy gradient updating mechanism is further introduced to form a learnable parameter field, collaborative optimization of a grid structure and parameter evolution is achieved through space-time alignment and topology reconstruction, and therefore the adaptability and dynamic updating performance of the model to a complex geological structure are remarkably enhanced. The finally output geologic model not only can accurately reflect the spatial relevance among the multi-source data of the mining area, but also has good continuity and interpretability, and effectively solves the problem that the multi-source heterogeneous geospatial data fusion modeling is difficult.

Description

Technical Field

[0001] This invention belongs to the field of mining technology, specifically relating to a method for optimizing geological parameters in open-pit mines based on deep reinforcement learning. Background Technology

[0002] In the geological modeling process of open-pit mines, effectively integrating multi-source heterogeneous geospatial data (such as topography, borehole data, remote sensing data, and geophysical data) and establishing a high-precision, dynamically updatable geological parameter model based on this data is a key challenge in current research and engineering practice. Traditional methods typically rely on manual experience to divide geological units and generate parameter distributions through static interpolation, which is difficult to adapt to the complex geological conditions and continuously changing mining environment of the mining area.

[0003] In existing technologies, a common approach is to process various types of data separately and then map them to a unified coordinate system before performing attribute assignment and interpolation calculations based on a fixed grid. While this method achieves data integration to some extent, it lacks the ability to adaptively extract spatial features at different scales and a dynamic adjustment mechanism for the matching relationship between historical observation data and the current parameter distribution. This results in significant deficiencies in boundary characterization, attribute continuity, and prediction accuracy in the constructed model. Furthermore, the fixed grid structure in traditional modeling processes prevents dynamic adjustment of topological relationships based on parameter evolution, further limiting the model's expressive power and update efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a method for optimizing geological parameters in open-pit mines based on deep reinforcement learning. This method can not only accurately reflect the spatial correlation between multi-source data in the mining area, but also has good continuity and interpretability, effectively solving the problem of difficulty in modeling the fusion of multi-source heterogeneous geospatial data.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for optimizing and modeling geological parameters in open-pit mines based on deep reinforcement learning, comprising the following steps:

[0006] A three-dimensional mesh framework is constructed to map multi-source geospatial data to a unified coordinate system. A multi-scale feature extractor is embedded in the three-dimensional mesh to obtain spatial attributes at different granularities.

[0007] An initial parameter distribution is generated based on the spatial attributes, the association between geological attributes and grid nodes is established, a dynamic weight allocation function is designed, and the weight coefficients in the dynamic weight allocation function are adjusted according to the matching degree between the initial parameter distribution and historical data.

[0008] The initial parameter distribution characteristics are updated by policy gradient to form a learnable parameter field. The parameter field is then spatiotemporally aligned with dynamic geological constraints, and the topological connectivity of the 3D mesh is corrected based on the alignment results.

[0009] The propagation path of the spatial attributes in the modified topology is recalculated, the interaction response between the propagation path and the weight coefficient is iteratively optimized, and the final geological model is output.

[0010] Preferably, mapping multi-source geospatial data to a unified coordinate system includes:

[0011] The original geographic information of the mining area is obtained, and the topographic elevation, rock strata distribution and geological fault points are extracted as the basic positioning basis. Based on the basic positioning basis, reference benchmarks are set, and a local coordinate system with the mining center as the origin is established. Coordinate transformation operations are applied to each type of geospatial data to convert it to the corresponding position in the local coordinate system. Non-uniform three-dimensional grid units are divided according to the density distribution of the converted data to form a discretized grid framework covering the entire mining area.

[0012] Preferably, a multi-scale feature extractor is embedded in the three-dimensional mesh to obtain spatial attributes at different granularities, including:

[0013] A neighborhood search range is defined for each node in the 3D mesh, and the local structural complexity is determined based on the attribute differences between adjacent nodes. A sliding sensing window is set based on the local structural complexity, and geometric shape and attribute change trends are extracted at different spatial levels. Directional sensitivity analysis is performed on the data distribution within the sliding sensing window to identify the main spatial change direction. The information extracted from each level is weighted and fused according to the main spatial change direction to generate a spatial attribute description with hierarchical differentiation.

[0014] Preferably, generating an initial parameter distribution based on the spatial attributes and establishing the association between geological attributes and grid nodes includes:

[0015] For each type of geological attribute, a corresponding response intensity threshold is set, and the response region is divided according to the change gradient of the spatial attribute. Within the response region, the spatial correspondence between the geological attribute and the neighboring grid nodes is matched to construct an attribute-node association table. The type of the node's dominant attribute is determined according to the co-occurrence of multiple attributes in the association table and marked as the node's core feature label. The node is subjected to attribute diffusion operation according to the core feature label to extend the local features to the adjacent regions and form a continuously distributed initial parameter field.

[0016] Preferably, a dynamic weight allocation function is designed, which adjusts the weight coefficients in the dynamic weight allocation function according to the matching degree between the initial parameter distribution and historical data, including:

[0017] Local features of the initial parameters distributed in the key change interval are extracted and matched with the corresponding historical records in time series. Based on the time series matching results, error-sensitive levels are divided, and weight response intensity benchmarks are set at different levels. The coefficients of each level are dynamically offset and corrected according to the weight response intensity benchmarks to generate a weight configuration that adapts to the current matching state. The weight configuration is integrated into a multi-dimensional allocation structure to form a weight output mode that adaptively adjusts with the evolution of spatial attributes.

[0018] Preferably, the initial parameter distribution characteristics are updated through policy gradients to form a learnable parameter field, including:

[0019] The parameter update direction is determined based on the current weight allocation result, and a gradient action region is delineated on the initial parameter distribution. The attribute change rate is calculated within the gradient action region to generate a local adjustment signal reflecting spatial evolution. The node parameters in the distribution are incrementally corrected according to the local adjustment signal to gradually improve the continuity of the parameter field. The corrected node parameters are incorporated into the global parameter structure to construct a three-dimensional parameter field with adaptive characteristics.

[0020] Preferably, aligning the parameter field with dynamic geological constraints in a spatiotemporal manner includes:

[0021] Extract the dominant attribute change trend in the parameter field and identify the corresponding geological evolution stage; determine the current dynamic constraint window based on the geological evolution stage and the real-time monitored geological activity time series; calculate the spatial offset within the dynamic constraint window and perform position correction on the region in the parameter field that exceeds the constraint range; adjust the attribute transition relationship within the parameter field according to the position correction result to achieve a continuous alignment with the geological constraints.

[0022] Preferably, the topological connectivity of the 3D mesh is corrected based on the alignment result, including:

[0023] Analyze the attribute misalignment regions in the spatiotemporal alignment results to identify node pairs with unstable connections; adjust the connection weights based on the spatial relative positions of the node pairs to redistribute the association strength between adjacent nodes; update the local mesh structure according to the adjusted connection weights to form a topology configuration that adapts to the current parameter distribution; extend the topology configuration to adjacent regions to complete the dynamic reconstruction of the overall mesh connection relationship.

[0024] Preferably, recalculating the propagation path of the spatial attribute in the modified topology includes:

[0025] Based on the corrected grid topology, the connectivity between nodes is determined, and a dynamic adjacency graph is constructed. In the dynamic adjacency graph, the attribute transmission direction is tracked, and the dominant propagation channel and its branch paths are defined. The attribute migration sequence is updated according to the flow distribution of the dominant propagation channel, and the information receiving order of each node is adjusted. The information receiving order is combined with the attribute change rate to generate a propagation path configuration that adapts to the current topology.

[0026] Preferably, the interaction response between the propagation path and the weight coefficients is iteratively optimized to output the final geological model, including:

[0027] Establish a dynamic coupling relationship between the propagation path and the weight coefficient, and record the attribute adjustment magnitude in each interaction; based on the attribute adjustment magnitude, select high-impact path-weight combinations and prioritize optimizing their collaborative response efficiency; update the global parameter transfer mode based on the collaborative response optimization and converge to a stable state; map the parameter distribution in the stable state to a three-dimensional geological entity structure to generate an interpretable final geological model.

[0028] Technical effects and advantages of the present invention: The open-pit mine geological parameter optimization modeling method based on deep reinforcement learning proposed in this invention has the following advantages compared with the prior art:

[0029] This invention achieves a unified spatial representation of multi-source data by constructing a three-dimensional non-uniform grid framework. It acquires hierarchically discriminative spatial attribute descriptions based on an embedded multi-scale feature extractor, and enhances the consistency between the initial parameter distribution and historical observation data by incorporating a dynamic weight allocation mechanism. The method also introduces a policy gradient update mechanism to form a learnable parameter field, and achieves synergistic optimization of grid structure and parameter evolution through spatiotemporal alignment and topology reconstruction, thereby significantly enhancing the model's adaptability to complex geological structures and its dynamic update performance. The final geological model not only accurately reflects the spatial correlation between multi-source data in the mining area but also possesses good continuity and interpretability, effectively solving the problem of difficult multi-source heterogeneous geospatial data fusion modeling. Attached Figure Description

[0030] Figure 1 This is a flowchart of the method for optimizing geological parameters in open-pit mines based on deep reinforcement learning, as described in this invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The specific embodiments described herein are merely used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] This invention provides, for example Figure 1 The open-pit mine geological parameter optimization modeling method based on deep reinforcement learning shown includes the following steps:

[0033] Step 1: Construct a 3D mesh framework to map multi-source geospatial data to a unified coordinate system; specifically including:

[0034] First, it is necessary to obtain the original geographical information of the mining area, including basic positioning data such as topographic elevation, rock strata distribution, and geological fault points. Topographic elevation data reflects the surface undulations of the mining area, rock strata distribution describes the spatial distribution of different underground lithologies, and geological fault points record the locations of geological fractures caused by tectonic activity. These data are acquired through methods such as ground mapping, remote sensing image interpretation, and borehole data analysis, which together constitute the basic geological framework of the mining area. Based on this information, a reasonable reference benchmark can be determined, and a local coordinate system can be established based on this benchmark.

[0035] In establishing a local coordinate system, the mining center is typically used as the origin to better match the actual needs of mining engineering. The coordinate system should consider the overall geometry of the mining area to ensure that all subsequent data transformations can be performed within the same reference system. For example, a Cartesian coordinate system can be used, where the x-axis and y-axis represent two orthogonal dimensions in the horizontal direction, and the z-axis represents height or depth in the vertical direction. This allows all geospatial data to be accurately mapped to this coordinate system, facilitating subsequent data fusion and modeling analysis.

[0036] Next, coordinate transformation operations need to be performed on various geospatial data to ensure that they accurately correspond to their respective positions in the local coordinate system. Since different data sources use different coordinate systems (such as GPS coordinates, national geodetic coordinate systems, etc.), corresponding coordinate transformations are necessary. This process typically relies on registration using known control points to ensure that all data are aligned within a unified coordinate framework. For example, least squares methods or affine transformations can be used to eliminate deviations between different coordinate systems, ensuring the geometric consistency of the data.

[0037] After completing the coordinate transformation, the next step is to create a non-uniform 3D mesh based on the transformed data density distribution. Traditional uniform mesh generation methods suffer from insufficient resolution or wasted computational resources when dealing with complex geological structures, while non-uniform mesh generation can dynamically adjust the mesh size according to the density of the data distribution. For example, smaller mesh cells are used in data-dense areas to improve modeling accuracy, while larger mesh cells are used in data-sparse areas to reduce computational burden. This mesh generation method not only improves data processing efficiency but also better adapts to the spatial heterogeneity of the geological structure in mining areas.

[0038] Ultimately, the discretized grid framework composed of non-uniform 3D grid cells can completely cover the entire mining area, providing a solid foundation for subsequent spatial attribute analysis and parameter optimization. This grid framework not only carries the mapping results of multi-source geospatial data but also provides the necessary computational environment for subsequent steps such as multi-scale feature extraction, dynamic weight allocation, and parameter field optimization. Through this series of steps, the geological information of the mining area is efficiently integrated, laying a good data foundation for further deep reinforcement learning modeling.

[0039] Step 2: Embed a multi-scale feature extractor in the 3D mesh to obtain spatial attributes at different granularities; specifically including:

[0040] First, a neighborhood search range is defined for each node in the 3D mesh to measure the attribute differences between adjacent nodes. Let a certain mesh node be... The set of nodes in its neighborhood is ,in Represents a node and The Euclidean distance between nodes is defined by r, which is a preset search radius. By calculating the attribute differences among nodes within this radius, the complexity of the local structure can be quantified. For example, if the attribute values ​​of nodes within a certain region fluctuate significantly, it indicates that the geological features of that region change drastically, requiring more refined feature extraction.

[0041] Secondly, a sliding sensing window is set based on the local structural complexity, and the geometric shape and attribute change trends are extracted at different spatial levels. The size of the sliding sensing window... Local structural complexity Decision, that is ,in It is a monotonically increasing function. For example, when When the value is high, choose the smaller value. To capture subtle geometric changes; when When it is lower, it expands. To obtain broader attribute trends, at each level, the attribute mean μ and standard deviation σ of the nodes within the window are calculated to describe the statistical characteristics of that region.

[0042] Subsequently, directional sensitivity analysis is performed on the data distribution within the sliding sensing window to identify the direction of change in the master control space. Assuming that attribute changes in a certain region mainly occur along a certain direction d, the intensity of variation in each direction can be evaluated by calculating the covariance matrix Σ. Let... The largest eigenvector of Σ corresponds to the direction of the dominant change. This method can distinguish the main geological trends in different regions, such as the extension direction of fault zones or the dip angle of rock strata.

[0043] Finally, the information extracted from each level is weighted and fused according to the direction of change in the master space to generate a spatial attribute description with hierarchical differentiation. Let the feature vector of the k-th level be... The corresponding weighting coefficient is Then the comprehensive eigenvector F is expressed as , where Σ represents the summation operation. Weight coefficients. Adjustments can be made based on the importance of the characteristics of each level; for example, levels with significant geological changes can be assigned higher importance. The value is adjusted to enhance its impact on the overall feature description.

[0044] Through the above steps, the multi-scale feature extractor can capture key attributes of the geological structure of the mining area at different spatial granularities, providing rich feature inputs for subsequent parameter optimization and modeling.

[0045] Step 3: Generate initial parameter distribution based on spatial attributes and establish the association between geological attributes and grid nodes; specifically including:

[0046] First, for each type of geological attribute, a corresponding response intensity threshold needs to be set to screen out areas with significant impact. Let the geological attributes be... The response intensity threshold is Then only if the spatial attribute of a certain grid node i is satisfy Only then is the node considered... This is part of the response area. The selection of this threshold needs to be combined with the geological characteristics of the mining area; for example, for rock layer thickness attributes, a higher threshold can be set. To avoid noise interference, the value can be appropriately reduced for fault density attributes. This is to capture more potential areas of geological anomalies.

[0047] Secondly, after determining the response area, it is necessary to match the spatial correspondence between geological attributes and adjacent grid nodes, and construct an attribute-node association table. Let the response area be... Contains several sets of grid nodes Then each node With attributes The strength of the association between them can be determined by their spatial attribute values. and The ratio determines, that is This weight Reflects the nodes For attributes The higher the weight of a node's response, the greater its influence on that attribute. By traversing all geological attributes and their corresponding response regions, a complete attribute-node association table can be generated for subsequent parameter optimization.

[0048] Furthermore, the dominant attribute type of a node is determined based on the co-occurrence of multiple attributes in the association table, and this type is marked as the core feature label. For a given grid node... Its related geological attribute set is The corresponding weights are respectively If a certain attribute weight If it is the largest among all attributes, then Marked as a node Core feature labels This label indicates the main geological features of the node, such as rock strata type, tectonic deformation intensity, or ore body enrichment. In this way, the dominant geological attributes of each node can be clearly identified in the 3D mesh, providing crucial reference for subsequent parameter optimization.

[0049] Finally, attribute diffusion is performed on the nodes according to the core feature labels, extending local features to adjacent regions to form a continuously distributed initial parameter field. Assume a certain node... The core feature labels are Then its surrounding adjacent nodes Will be affected The influence, the intensity of the influence, is determined by the spatial distance between the two. The decision is made. The diffusion formula is expressed as: ,in, Represents a node For attributes The initial parameter values, where λ is the attenuation coefficient. For nodes and The distance between them. This formula can be used to calculate the distance between adjacent nodes. parameter values This process is gradually extended to the entire mining area, forming a continuous and reasonable distribution of initial parameters.

[0050] Through the above steps, the correlation between geological attributes and grid nodes can be effectively established, and a preliminary parameter field can be generated, providing a reliable foundation for subsequent dynamic weight adjustment and parameter optimization.

[0051] Step 4: Design a dynamic weight allocation function, and adjust the weight coefficients in the dynamic weight allocation function according to the matching degree between the initial parameter distribution and historical data; specifically including:

[0052] First, it is necessary to extract local features of key variation intervals from the initial parameter distribution and perform time-series matching with the corresponding historical data. Assume the initial parameter field P(x,y,z) has a value at a certain spatial location (x,y,z) of... The historical data H(x,y,z) records the values ​​at the same positions as... The local error between the two Represented as: To measure the relative importance of the error, a normalization factor ε can be introduced, allowing the error to vary between 0 and 1. Where max(E) represents the maximum value of all errors, and ε is a small constant to prevent division by zero. This is achieved by calculating the values ​​at each spatial location. It can identify areas with high errors and classify error-sensitive levels accordingly.

[0053] Secondly, a benchmark for the weighted response strength is set based on the error sensitivity level. Assume there are L levels of error sensitivity, and the benchmark weight for each level l is... Then the weighting coefficient Can be determined by the level Decide: ;in, The basis for the division is The size. For example, if ≤0.2, then =1 indicates a low error-sensitive level; if 0.2 < If ≤0.5, then l_i=2, indicating a medium error-sensitive level; if >0.5, then =3 indicates a high error-sensitive level. In this way, regions with larger errors can be given higher weight adjustment priority, thus converging to a better parameter distribution more quickly.

[0054] Furthermore, the coefficients at each level are dynamically offset and corrected based on the weight response strength benchmark to generate a weight configuration adapted to the current matching state. Considering that the error changes over time, a time decay factor τ is introduced to allow the weight coefficients to adjust according to the error's changing trend. New weight coefficients... Represented as: ;

[0055] Where δ is the learning rate. This is the error from the previous period. This represents the error for the current time period. If... < This indicates that the error is decreasing. It will be appropriately lowered to reduce unnecessary adjustments; conversely, if t> v, then This will increase in speed to accelerate parameter correction.

[0056] Finally, the adjusted weight configuration is integrated into the multidimensional allocation structure, forming a weight output mode that adaptively adjusts as spatial attributes evolve. This structure is represented as: ;

[0057] in, The initial weights for each dimension of the feature are: These are the dynamically adjusted weight coefficients. This formula ensures that the weights can be automatically adjusted according to error changes at different spatial locations, thereby improving the model's adaptability and stability.

[0058] Through the above steps, the dynamic weight allocation function can effectively adjust the parameter influence factors of each region, enabling the model to have stronger adaptability when facing different geological conditions, thereby improving the accuracy and reliability of geological modeling.

[0059] Step 5: Update the initial parameter distribution characteristics through policy gradients to form a learnable parameter field; specifically including:

[0060] First, the parameter update direction is determined based on the current weight allocation result. Assume the parameter field P(x,y,z) has a value at a certain spatial location (x,y,z) of... The corresponding weighting coefficient is Then the parameter update direction The direction of the gradient, guided by the weights, can be determined. Specifically, Represented as: Where α is the learning rate. for The gradient estimate at point . Calculations can be performed based on the parameter changes of surrounding nodes, for example, using the finite difference method: ,in, These are the parameter values ​​of neighboring nodes. Let be the distance between nodes i and j. This formula can be used to determine the trend of parameter field changes in a local region and, based on this, determine the direction of parameter updates.

[0061] Secondly, the gradient action region is defined within the parameter field to limit the range of parameter updates. The choice of the gradient action region depends on the weighting coefficients. The region with the highest weight should be prioritized for parameter adjustment. Let the weight threshold for a certain region be... Then the gradient action region Represented as: Within this region, parameter updates will be activated, while nodes outside the region will remain unchanged. This ensures that parameter updates are concentrated in critical areas with large errors, improving optimization efficiency.

[0062] Furthermore, the rate of change of attributes within the gradient's effective region is calculated to generate a local adjustment signal reflecting spatial evolution. (Attribute change rate) It can be used to measure the evolution trend of the parameter field in a local region, and its calculation formula is as follows: ,in, and These represent the parameter values ​​at the current and previous times, respectively, with Δt being the time step. If... A larger value indicates that the geological characteristics of the area are changing drastically, requiring more frequent parameter updates; conversely, a smaller value indicates that the area is relatively stable, and the update frequency can be appropriately reduced.

[0063] Finally, the node parameters in the distribution are incrementally corrected based on the local adjustment signal, gradually improving the continuity of the parameter field. The update formula for incremental correction is expressed as: Where β is a correction coefficient used to control the update magnitude. This formula allows for small adjustments to node parameters in each iteration, avoiding instability in the parameter field caused by a large, one-time modification. As the number of iterations increases, the parameter field gradually approaches the optimal solution, forming a three-dimensional parameter field with adaptive properties.

[0064] Through the above steps, the policy gradient update method can effectively optimize the initial parameter distribution, making it more adaptable to different geological conditions, thereby improving the accuracy and stability of geological modeling.

[0065] Step Six: Spatiotemporally align the parameter field with dynamic geological constraints; specifically including:

[0066] First, it is necessary to extract the dominant attribute variation trend from the parameter field and identify its corresponding geological evolution stage. Let the attribute variation rate of the parameter field P(x,y,z) at a certain spatial location (x,y,z) be R(x,y,z). Then, the dominant attribute variation trend can be determined by the direction of the maximum variation rate. For example, if the gradient of R(x,y,z) in a certain region is the largest in the x-direction, it indicates that the geological evolution of that region mainly proceeds along the x-direction. Furthermore, the variation rate sequence R(t) can be matched with known geological evolution stages to determine the current evolution stage. For example, if R(t) shows an upward trend, it corresponds to an active tectonic period; if R(t) tends to stabilize, it indicates a geologically stable period.

[0067] Secondly, based on the geological evolution stages matched with real-time monitored geological activity time series, the current dynamic constraint window is determined. Let the geological evolution stage S be labeled by historical data, and the real-time monitored time series M(t) record the changes in current geological activity. Then, the dynamic constraint window... It can be determined by a similarity metric. For example, the correlation coefficient ρ(S,M(t)) can be used to measure the degree of matching. At this point, it is considered that the current stage is within the constraint window of stage S. Here, θ is a preset matching threshold. This method allows for dynamic adjustment of the evolution constraints of the parameter field, ensuring it always conforms to the latest geological activity characteristics.

[0068] Furthermore, the spatial offset is calculated within the dynamic constraint window, and positional correction is performed on regions in the parameter field that exceed the constraint range. Suppose the parameter value P(x,y,z) in a certain region deviates from the expected value Q(x,y,z) within the constraint window; then the spatial offset ΔP(x,y,z) is expressed as: ;

[0069] If ΔP(x,y,z) exceeds the allowable error range ε, then the parameter field needs to be corrected. The correction formula is expressed as: Where λ is the correction coefficient used to control the adjustment range. This formula allows for the gradual reduction of the deviation between the parameter field and geological constraints, making it closer to actual geological conditions.

[0070] Finally, based on the location correction results, the attribute transition relationships within the parameter field are adjusted to achieve continuous alignment with geological constraints. Suppose that the parameter field in a certain region experiences abrupt attribute changes, and the parameter values ​​of its adjacent regions are respectively... and Therefore, the adjustment of attribute transition relationships can be accomplished through interpolation calculations. For example, using a linear interpolation method: Where x is the position of the point to be adjusted. and These are the boundary points of adjacent regions. This method can smooth the transition of properties in the parametric field, maintaining its spatial continuity and consistency with geological constraints.

[0071] Through the above steps, the parameter field can dynamically adapt to the geological evolution process and maintain a reasonable parameter distribution in different geological stages, thereby improving the accuracy and stability of geological modeling.

[0072] Step 7: Correct the topological connectivity of the 3D mesh based on the alignment results; specifically including:

[0073] First, it is necessary to analyze the attribute misalignment regions in the spatiotemporal alignment results and identify node pairs with unstable connections. Let the error between the parameter field P(x,y,z) and its expected value Q(x,y,z) in a certain region be E(x,y,z) = |P(x,y,z) - Q(x,y,z)|. If E(x,y,z) exceeds a preset error threshold ε, then the region is determined to be an attribute misalignment region. Within this region, the connection relationship between adjacent nodes no longer conforms to the spatial evolution trend of the parameter field and needs adjustment. For example, if the attribute differences between node i and node j are large, but they still maintain a strong connection weight, it leads to parameter propagation instability, thus affecting the accuracy of the model.

[0074] Secondly, the connection weights are adjusted based on the spatial relative positions of the node pairs to redistribute the association strength between adjacent nodes. Let the initial connection weight between node i and node j be... Its adjusted weight This can be determined by the differences in their attributes. and spatial distance Joint decision: ;

[0075] in, = This represents the attribute difference between nodes i and j, where λ is the attenuation coefficient used to control the impact of attribute differences on connection weights. If If it is larger, then This will decrease accordingly, weakening the connection strength between nodes with significantly different attributes; conversely, if Smaller, then Maintain a high value to preserve a stable parameter transfer relationship.

[0076] Furthermore, the local mesh structure is updated based on the adjusted connection weights to form a topology configuration adapted to the current parameter distribution. Let the set of nodes in a certain local region be denoted as . Its corresponding connection weight matrix is The new topology can then be determined by the Maximum Spanning Tree (MST) algorithm. The goal of MST is to maximize the overall connection weights while maintaining connectivity, thus ensuring that the mesh structure can adapt to changes in the parameter field while maintaining high stability. Specifically, the MST can be constructed using either the Kruskal algorithm or the Prim algorithm, selecting the edge with the highest weight to connect all nodes, forming the optimal local topology configuration.

[0077] Finally, the local topology configuration is extended to adjacent regions, completing the dynamic reconstruction of the overall mesh connectivity. Let the local topology configuration be... It has been determined that its adjacent areas The set of nodes is The expansion process can then be performed in the following ways:

[0078] for Each node in Calculate its relationship with Attribute similarity of all nodes And select the node with the highest similarity. Establish a connection.

[0079] Update the global connection weight matrix based on the newly established connections. This ensures the connectivity of the overall mesh and the consistency of the parameter field.

[0080] Through the above steps, the topological connectivity of the 3D mesh can be dynamically adjusted to adapt to the changing trends of the parameter field, thereby improving the adaptability and stability of geological modeling.

[0081] Step 8: Recalculate the propagation path of spatial properties in the modified topology; specifically including:

[0082] First, based on the corrected grid topology, the connectivity between nodes is determined, and a dynamic adjacency graph is constructed. Let the set of adjacent nodes of a given node i be denoted as . The connection weight between nodes i and j This determines its connectivity. If If the adjacency value is greater than or equal to τ (where τ is a preset connectivity threshold), then nodes i and j are considered to have a valid connection; otherwise, they are considered disconnected. A dynamic adjacency matrix can be constructed by traversing the adjacency relationships of all nodes. ,in = , is used to describe the connectivity state of the entire grid.

[0083] Secondly, the attribute propagation direction is tracked in the dynamic adjacency graph to delineate the dominant propagation channel and its branch paths. Suppose an attribute propagates from the starting node s; its propagation path can be determined by a shortest path algorithm (such as Dijkstra's algorithm). For any node i, the shortest path length from i to s is... It can be calculated using the following formula: ,in, The propagation cost between nodes i and j is typically represented by the connection weights. Decision, that is ε is a small constant to prevent division by zero. Using this formula, the propagation distance of all nodes relative to s can be calculated, and the main propagation path of attribute propagation can be determined accordingly. The dominant propagation path is the path extending from s along the path with the minimum propagation cost, while branch paths are secondary propagation routes that branch off from the main path.

[0084] Furthermore, the attribute migration sequence is updated according to the flow distribution of the dominant propagation channel, adjusting the information reception order of each node. Let the set of upstream nodes of a certain node i be denoted as . Its downstream node set is Then the attribute migration sequence T can be determined by a topological sorting algorithm. For example, using the breadth-first search (BFS) method, starting from s, all reachable nodes are visited sequentially and arranged in the order of visit to form the attribute migration sequence. This sequence ensures that attributes flow through the network in the correct propagation order, avoiding information conflicts or delays.

[0085] Finally, the information reception order is combined with the attribute change rate to generate a propagation path configuration adapted to the current topology. Let the attribute change rate of a certain node i be... The order in which it receives information is as follows Then its final attribute update formula is expressed as: Where α is the learning rate. The weight for the information receiving order of node j. Let be the attribute value of node j. This formula ensures that the attribute propagation process follows both the connectivity of the network topology and the rate of attribute change of different nodes, thereby generating a more accurate propagation path configuration.

[0086] Step Nine: Iteratively optimize the interaction response between the propagation path and the weighting coefficients to output the final geological model; specifically including:

[0087] First, establish a dynamic coupling relationship between the propagation path and the weight coefficients, and record the attribute adjustment magnitude in each interaction. Let a certain propagation path... The weighting coefficient is The corresponding attribute adjustment range It can be calculated using the following formula: ;

[0088] in, and These represent the path i and i respectively. The change in attribute values ​​before and after uppropagation. This formula quantifies the impact of each propagation path on attribute adjustment and establishes the coupling relationship between propagation paths and weight coefficients. For example, if... A larger value indicates that the path has a strong influence on the evolution of the parameter field and should be given higher priority in subsequent optimization processes.

[0089] Secondly, based on the attribute adjustment magnitude, high-impact path-weight combinations are selected, and their collaborative response efficiency is optimized first. Let the set of adjustment magnitudes for all paths be... Then, high-impact paths can be composed of paths where ΔA is greater than a certain threshold θ: For each path in H The corresponding weighting coefficient The following optimization objectives can be adjusted: Where β is the adjustment coefficient, and max(ΔA) is the maximum adjustment magnitude among all paths. This formula can enhance the weight of high-influence paths, enabling them to play a greater role in subsequent propagation and thus accelerating the optimization process of the parameter field.

[0090] Furthermore, based on the cooperative response optimization, the global parameter transfer mode is updated and converged to a stable state. Let the global parameter field at a certain time step t be... Its updated parameter field It can be calculated using the following formula: ;

[0091] Where γ is the global learning rate. This represents the combined impact of all paths on node i. Through continuous iterative updates, the parameter field will gradually approach a stable state until the parameter changes in adjacent time steps are less than a certain convergence threshold ε. At this point, the model reaches convergence, and the evolution of the parameter field tends to stabilize, indicating that the interaction between the propagation path and the weight coefficients has reached the optimal balance.

[0092] Finally, the parameter distribution under steady-state conditions is mapped to a three-dimensional geological entity structure, generating an interpretable final geological model. Let the distribution of a certain geological attribute A in the parameter field be... Then the final geological model M(x,y,z) can be constructed by the following formula: ;

[0093] in, This is the visualization weighting coefficient for attribute A, used to adjust the display proportion of different geological attributes in the final model. This formula maps numerical values ​​in the parameter field to specific geological entities, such as rock strata, faults, or ore body distributions, thereby generating an intuitive and interpretable 3D geological model.

[0094] This invention achieves a unified spatial representation of multi-source data by constructing a three-dimensional non-uniform grid framework. It obtains spatial attribute descriptions with hierarchical discrimination based on the embedded multi-scale feature extractor, and improves the consistency between the initial parameter distribution and historical observation data by combining a dynamic weight allocation mechanism.

[0095] This method also introduces a policy gradient update mechanism to form a learnable parameter field, and achieves synergistic optimization of grid structure and parameter evolution through spatiotemporal alignment and topology reconstruction, thereby significantly enhancing the model's adaptability to complex geological structures and its dynamic update performance. The final output geological model not only accurately reflects the spatial correlation between multi-source data in the mining area, but also possesses good continuity and interpretability, effectively solving the problem of difficult modeling of multi-source heterogeneous geospatial data fusion.

[0096] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing geological parameters in open-pit mines based on deep reinforcement learning, characterized in that, Includes the following steps: A three-dimensional mesh framework is constructed to map multi-source geospatial data to a unified coordinate system. A multi-scale feature extractor is embedded in the three-dimensional mesh to obtain spatial attributes at different granularities. An initial parameter distribution is generated based on the spatial attributes, the association between geological attributes and grid nodes is established, a dynamic weight allocation function is designed, and the weight coefficients in the dynamic weight allocation function are adjusted according to the matching degree between the initial parameter distribution and historical data. The initial parameter distribution characteristics are updated by policy gradient to form a learnable parameter field. The parameter field is then spatiotemporally aligned with dynamic geological constraints, and the topological connectivity of the 3D mesh is corrected based on the alignment results. The propagation path of the spatial attributes in the modified topology is recalculated, the interaction response between the propagation path and the weight coefficient is iteratively optimized, and the final geological model is output.

2. The method for optimizing geological parameters in open-pit mines based on deep reinforcement learning according to claim 1, characterized in that: Mapping multi-source geospatial data to a unified coordinate system includes: The original geographic information of the mining area is obtained, and the topographic elevation, rock strata distribution and geological fault points are extracted as the basic positioning basis. Based on the basic positioning basis, reference benchmarks are set, and a local coordinate system with the mining center as the origin is established. Coordinate transformation operations are applied to each type of geospatial data to convert it to the corresponding position in the local coordinate system. Non-uniform three-dimensional grid units are divided according to the density distribution of the converted data to form a discretized grid framework covering the entire mining area.

3. The method for optimizing geological parameters in open-pit mines based on deep reinforcement learning according to claim 1, characterized in that: A multi-scale feature extractor is embedded in the three-dimensional mesh to obtain spatial properties at different granularities, including: A neighborhood search range is defined for each node in the 3D mesh, and the local structural complexity is determined based on the attribute differences between adjacent nodes. A sliding sensing window is set based on the local structural complexity, and geometric shape and attribute change trends are extracted at different spatial levels. Directional sensitivity analysis is performed on the data distribution within the sliding sensing window to identify the main spatial change direction. The information extracted from each level is weighted and fused according to the main spatial change direction to generate a spatial attribute description with hierarchical differentiation.

4. The method for optimizing geological parameters in open-pit mines based on deep reinforcement learning according to claim 1, characterized in that: Based on the aforementioned spatial attributes, an initial parameter distribution is generated, and the association between geological attributes and grid nodes is established, including: For each type of geological attribute, a corresponding response intensity threshold is set, and the response region is divided according to the change gradient of the spatial attribute. Within the response region, the spatial correspondence between the geological attribute and the neighboring grid nodes is matched to construct an attribute-node association table. The type of the node's dominant attribute is determined according to the co-occurrence of multiple attributes in the association table and marked as the node's core feature label. The node is subjected to attribute diffusion operation according to the core feature label to extend the local features to the adjacent regions and form a continuously distributed initial parameter field.

5. The method for optimizing geological parameters in open-pit mines based on deep reinforcement learning according to claim 1, characterized in that: Design a dynamic weight allocation function, and adjust the weight coefficients in the dynamic weight allocation function according to the matching degree between the initial parameter distribution and historical data, including: Local features of the initial parameters distributed in the key change interval are extracted and matched with the corresponding historical records in time series. Based on the time series matching results, error-sensitive levels are divided, and weight response intensity benchmarks are set at different levels. The coefficients of each level are dynamically offset and corrected according to the weight response intensity benchmarks to generate a weight configuration that adapts to the current matching state. The weight configuration is integrated into a multi-dimensional allocation structure to form a weight output mode that adaptively adjusts with the evolution of spatial attributes.

6. The method for optimizing geological parameters in open-pit mines based on deep reinforcement learning according to claim 1, characterized in that: The initial parameter distribution characteristics are updated through policy gradients to form a learnable parameter field, including: The parameter update direction is determined based on the current weight allocation result, and a gradient action region is delineated on the initial parameter distribution. The attribute change rate is calculated within the gradient action region to generate a local adjustment signal reflecting spatial evolution. The node parameters in the distribution are incrementally corrected according to the local adjustment signal to gradually improve the continuity of the parameter field. The corrected node parameters are incorporated into the global parameter structure to construct a three-dimensional parameter field with adaptive characteristics.

7. The method for optimizing geological parameters in open-pit mines based on deep reinforcement learning according to claim 1, characterized in that: The parameter field is spatiotemporally aligned with dynamic geological constraints, including: Extract the dominant attribute change trend in the parameter field and identify the corresponding geological evolution stage; determine the current dynamic constraint window based on the geological evolution stage and the real-time monitored geological activity time series; calculate the spatial offset within the dynamic constraint window and perform position correction on the region in the parameter field that exceeds the constraint range; adjust the attribute transition relationship within the parameter field according to the position correction result to achieve a continuous alignment with the geological constraints.

8. The method for optimizing geological parameters in open-pit mines based on deep reinforcement learning according to claim 1, characterized in that: Correcting the topological connectivity of the 3D mesh based on the alignment results includes: Analyze the attribute misalignment regions in the spatiotemporal alignment results to identify node pairs with unstable connections; adjust the connection weights based on the spatial relative positions of the node pairs to redistribute the association strength between adjacent nodes; update the local mesh structure according to the adjusted connection weights to form a topology configuration that adapts to the current parameter distribution; extend the topology configuration to adjacent regions to complete the dynamic reconstruction of the overall mesh connection relationship.

9. The method for optimizing geological parameters in open-pit mines based on deep reinforcement learning according to claim 8, characterized in that: Recalculating the propagation path of the spatial properties in the modified topology includes: Based on the corrected grid topology, the connectivity between nodes is determined, and a dynamic adjacency graph is constructed. In the dynamic adjacency graph, the attribute transmission direction is tracked, and the dominant propagation channel and its branch paths are defined. The attribute migration sequence is updated according to the flow distribution of the dominant propagation channel, and the information receiving order of each node is adjusted. The information receiving order is combined with the attribute change rate to generate a propagation path configuration that adapts to the current topology.

10. The method for optimizing geological parameters in open-pit mines based on deep reinforcement learning according to claim 9, characterized in that: Iteratively optimize the interaction response between the propagation path and the weight coefficients to output the final geological model, including: Establish a dynamic coupling relationship between the propagation path and the weight coefficient, and record the attribute adjustment magnitude in each interaction; based on the attribute adjustment magnitude, select high-impact path-weight combinations and prioritize optimizing their collaborative response efficiency; update the global parameter transfer mode based on the collaborative response optimization and converge to a stable state; map the parameter distribution in the stable state to a three-dimensional geological entity structure to generate an interpretable final geological model.

Images