A geological exploration risk assessment method based on big data analysis

Abstract

The application discloses a geological exploration risk assessment method based on big data analysis, comprising the following steps: S1, organizing an exploration object data set and a historical geological risk event position mapping relationship; S2, constructing a geological risk factor matrix to generate a geological genesis atlas; S3, inputting the risk factor into an improved optimal parameter geographic detector algorithm to obtain risk factor stratification results and risk factor interpretation intensity; S4, analyzing the risk factor stratification results for a risk main control relationship to form a risk interpretation result; S5, rearranging risk factor codes to form a risk assessment characteristic sequence; S6, inputting the risk assessment characteristic sequence into an improved evidence neural network model to obtain an exploration object risk evidence vector; and S7, determining a risk grade and a risk uncertainty to backfill to form a geological exploration risk zoning result and an exploration priority result. The application can improve the geological exploration risk assessment accuracy, interpretability and priority ranking reliability.

Description

Technical Field

[0001] This invention relates to the field of intelligent risk assessment technology for geological exploration, and in particular to a geological exploration risk assessment method based on big data analysis. Background Technology

[0002] With the continuous advancement of deep mineral exploration, resource evaluation in complex tectonic zones, and engineering geological surveys, geological exploration activities face increasingly complex data types and more diverse sources of risk. Existing geological exploration risk assessment methods largely rely on expert experience-based weighting, single-factor overlay analysis, or conventional machine learning models. While these methods can identify risk factors such as fault structures, lithological fracturing, hydrological disturbances, and geophysical anomalies to a certain extent, they still have shortcomings in practical applications: geological data from different sources are difficult to organize uniformly according to the exploration object; there is a lack of effective correlation between historical geological risk events and risk factor stratification results; and risk factor combinations are prone to spurious correlation judgments that are statistically correlated but lack geological evidence. Furthermore, traditional models often directly output risk levels, failing to simultaneously provide the risk source and uncertainty, resulting in insufficient interpretability of the assessment results and difficulty in supporting supplementary exploration deployment and exploration priority decisions.

[0003] Therefore, how to provide a geological exploration risk assessment method based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a geological exploration risk assessment method based on big data analysis. This invention improves the optimal parameter geographic detector algorithm by incorporating an event projection correction function into the candidate stratification process, embedding the location mapping relationship of historical geological risk events into the stratification process. This ensures that the risk factor stratification results are determined not only by spatial differentiation interpretability but also by the actual distribution of historical geological risk events, thereby reducing stratification misjudgments caused by relying solely on statistical differentiation. Simultaneously, by introducing a geological genetic map through a target stratification function, combinations of risk factors without geological genetic continuity are restricted, ensuring that the identification of combined risk-causing factors has a clear causal basis and reducing the inclusion of pseudo-correlated combinations without causal support in the risk interpretation results. Furthermore, this invention sets up a stratified evidence embedding layer in the improved evidence neural network model, embedding the risk factor stratification results, risk factor interpretability, master control markers, and combination markers into the evidence generation process. This allows risk level, risk uncertainty, and risk source to be collaboratively output based on the same risk evidence vector, improving the accuracy, interpretability, and reliability of exploration priority ranking in geological exploration risk assessment results.

[0005] A geological exploration risk assessment method based on big data analysis according to an embodiment of the present invention includes the following steps: S1. Aggregate multi-source big data on geological exploration in the target exploration area, and organize it into a dataset of exploration objects and a mapping relationship between the locations of historical geological risk events according to the exploration objects; S2. Configure risk factor fields based on the exploration object dataset, construct a geological risk factor matrix, and generate a geological genetic map based on the geological genetic relationship between the risk factor fields. S3. Input the risk factors in the geological risk factor matrix into the improved optimal parameter geographic detector algorithm, construct the event projection correction function, embed the location mapping relationship of historical geological risk events into the candidate stratification process, construct the target stratification function, introduce the geological gene map to impose causal constraints on the combination of risk factors, and obtain the risk factor stratification results and the risk factor interpretation intensity. S4. Analyze the risk control relationship of the risk factor stratification results, identify the risk factors whose explanatory strength meets the control screening conditions as the main control risk factors, and identify the combination of risk factors along the genetic inheritance path in the geological genetic map to form the risk interpretation results. S5. Rearrange the risk factor codes based on the risk factor stratification results and risk interpretation results to form a risk assessment feature sequence; S6. Input the risk assessment feature sequence into the improved evidence neural network model containing a hierarchical evidence embedding layer to obtain the risk evidence vector of the exploration object. S7. Determine the risk level and risk uncertainty based on the risk evidence vector of the exploration object, determine the source of risk in combination with the risk interpretation results, and backfill to form the geological exploration risk zoning results and exploration priority results.

[0006] Optionally, S1 specifically includes: S11. Project the boundary of the target exploration area, the layout of exploration operations, and the existing exploration control positions onto a unified spatial reference, delineate the set of exploration objects, and assign object numbers, object boundaries, and object type markers to each exploration object; S12. Perform source identification, spatial attribution identification, and time period attribution identification on multi-source big data of geological exploration, and link the data that falls within the object boundary and the data that has an operational control relationship with the exploration object to the corresponding exploration object respectively; S13. Perform coverage segmentation and object attribution assignment on data spanning multiple exploration objects, and encapsulate the data that has completed the attribution assignment under the same exploration object within the object to obtain the exploration object dataset; S14. Determine the location and impact range of historical geological risk events, and link them to the corresponding exploration objects to form a location mapping relationship of historical geological risk events.

[0007] Optionally, S2 specifically includes: S21. Classify the data items in the exploration object dataset according to the causes of geological structure, stratigraphy, abnormal response, hydrological environment, engineering disturbance and historical events, and configure the risk factor field set. S22. Perform field repositioning on the data items in each exploration object, convert data items under the same causal category into corresponding risk factor field values, and configure data validity markers for missing fields. S23. Using the exploration object number as the matrix row index and the risk factor field as the matrix column index, the risk factor field values ​​and data validity markers are assembled in the same position to obtain the geological risk factor matrix. S24. The risk factor fields are processed into nodes according to their causal categories, and the geological causal relationship between the risk factor fields is converted into directed relationship edges to obtain the initial geological causal relationship diagram. S25. Verify the endpoints, directions, and duplicates of the directed edges in the initial geological genetic relationship diagram, retaining the directed edges with complete endpoints and consistent directions to obtain the geological genetic map.

[0008] Optionally, S3 specifically includes: S31. In the improved optimal parameter geographic detector algorithm, each risk factor field in the geological risk factor matrix is ​​expanded into a field value sequence, and candidate discrete segmentation is performed on the field value sequence to obtain a candidate stratification result set. S32. Perform interval assignment calibration on each candidate stratification result in the candidate stratification result set, label the stratification interval in each candidate stratification result as a candidate risk layer, and assign each exploration object to the corresponding candidate risk layer according to the corresponding risk factor value, and calculate the initial stratification score value of the candidate stratification result according to the factor detection rules of the geographic detector. S33. Project the location mapping relationship of historical geological risk events onto the candidate risk layer, count the density value of historical geological risk events in each candidate risk layer, and calculate the event density difference value between adjacent candidate risk layers. S34. Construct an event projection correction function. The event projection correction function takes the initial stratification score value, the historical geological risk event density value and the event density difference value as input. It performs score retention on the candidate stratification results where the historical geological risk events are concentrated in the high-risk layer, and performs score reduction on the candidate stratification results where the distribution of historical geological risk events is inconsistent with the candidate risk level. It outputs the event-corrected stratification score value. S35. Convert the geological genetic map into a risk factor causal adjacency table, configure effective causal adjacency values ​​for risk factor combinations with geological genetic inheritance relationships, and configure invalid causal adjacency values ​​for risk factor combinations without geological genetic inheritance relationships. S36. Construct a target stratification function. The target stratification function takes the event-corrected stratification score and the risk factor causal adjacency value as input. It performs masking on the candidate stratification results corresponding to the invalid causal adjacency values, retains the candidate stratification results corresponding to the valid causal adjacency values, and selects the target stratification result from the retained candidate stratification results. S37. The stratification boundary, stratification number, and score value corresponding to the target stratification result are determined as the risk factor stratification result and the risk factor explanatory strength.

[0009] Optionally, S4 specifically includes: S41. Assign intervals to each risk factor in the risk factor stratification results according to the stratification number, and assign the explanatory strength of the risk factor to the corresponding stratification interval to obtain the single-factor risk explanation record. S42. Sort the risk factors in the single-factor risk interpretation record according to their explanatory strength. Mark the risk factors whose explanatory strength meets the main control screening criteria and whose historical geological risk event density in the corresponding stratified interval meets the event concentration criteria as the main control risk factors. S43. Locate the causal path of the main control risk factors in the geological causal map, connect the main control risk factors with geological causal relationship to obtain the candidate risk factor combination. S44. Cross-over the risk factor stratification intervals in the candidate risk factor combination, classify the exploration objects falling into the same overlapping interval into the same combination risk unit, and calculate the combination explanation strength of the combination risk unit. S45. Compare the combined explanatory strength of the combined risk unit with the corresponding single-factor risk explanatory record, configure a combined risk-causing label for candidate risk-causing factors whose explanatory strength is enhanced after combination, and determine the combined risk-causing factors. S46. Encapsulate the main control risk factors, main control stratification intervals, combined risk factors, combined risk units, and geological origination paths into risk interpretation results.

[0010] Optionally, S5 specifically includes: S51. Match the stratification number in the risk factor stratification result with the main control risk factor and combined risk-causing factor in the risk interpretation result, and configure the main control label, combined label and stratification label for the risk factor field in the geological risk factor matrix; S52. Arrange the risk factor fields with the configured master control tags first, and determine the field order from high to low according to the explanatory strength of the risk factors; S53. Map the risk factor fields of the configured combination markers to the directed relation edges in the geological genealogy, extract the starting risk factor fields and the ending risk factor fields corresponding to the same combination of risk factors, and arrange them adjacently in the order of starting risk factor fields first and ending risk factor fields last to form a combination field segment. S54. Risk factor fields without configured master control and combination labels are retained in the supplementary field section and listed after the master control and combination field sections; S55. Encapsulate the risk factor values, stratification numbers, risk factor interpretation strengths, main control markers, and combination markers corresponding to each exploration object in the same position according to the rearranged field order to form a risk assessment feature sequence.

[0011] Optionally, the improved evidence neural network model specifically includes a risk feature embedding layer, a hierarchical evidence embedding layer, a risk latent representation mapping layer, and an evidence output layer: The risk feature embedding layer divides the risk assessment feature sequence into field feature units according to the order of risk factor fields. Each field feature unit includes risk factor code, hierarchical number, risk factor explanatory strength, master control label and combination label. The risk factor codes in the field feature units are expanded into one-dimensional factor code vectors, and the stratified number, main control label, and combined label are converted into stratified label vectors, main control label vectors, and combined label vectors, respectively. The one-dimensional factor encoding vector, hierarchical label vector, master control label vector and combined label vector are concatenated in a fixed order within the fields to obtain the field embedding vector, and the field embedding vectors are arranged in the order of the risk factor fields to form a field embedding sequence. The hierarchical evidence embedding layer performs field matching and hierarchical position matching between the field embedding vectors in the field embedding sequence and the risk factor hierarchical results. It then performs hierarchical anchoring and evidence response transformation on the matched field embedding vectors to generate a hierarchical evidence response sequence. The risk implicit representation mapping layer divides the hierarchical evidence response sequence into the main control evidence segment, the combined evidence segment, and the supplementary evidence segment according to the main control label and the combined label. The hierarchical evidence response vector in each evidence segment is accumulated dimension by dimension to obtain the main control evidence vector, the combined evidence vector, and the supplementary evidence vector. The main evidence vector, combined evidence vector, and supplementary evidence vector are concatenated in sequence and then processed by linear mapping to obtain the risk hidden representation of the exploration object; The evidence output layer maps the implicit risk representation of the exploration object to the evidence channels corresponding to low risk, medium risk, high risk and extremely high risk, and arranges the output values ​​of each evidence channel in order of risk level to obtain the risk evidence vector of the exploration object.

[0012] Optionally, the step of matching the field embedding vectors in the field embedding sequence with the risk factor stratification results for field matching and stratification position matching, and performing stratification anchoring and evidence response transformation on the matched field embedding vectors to generate a stratified evidence response sequence specifically includes: The risk factor stratification results are split into field stratification records according to the risk factor field. Each field stratification record includes the risk factor field, stratification number, stratification boundary, and risk factor explanatory strength. A stratification anchor vector is configured for each stratification number, and a stratification anchor table is established. For the target field embedding vector in the field embedding sequence, extract the corresponding risk factor field and stratification number, match the risk factor field and stratification number with the stratification anchoring table, and locate the target stratification anchoring vector. The target field embedding vector and the target hierarchical anchoring vector are added dimension by dimension to generate a hierarchical alignment vector; The explanatory strength of the risk factors corresponding to the target field embedding vector is expanded into an explanatory strength correction vector according to the dimensions of the hierarchical alignment vector. The hierarchical alignment vector and the explanatory strength correction vector are then multiplied dimension by dimension to generate a hierarchical response vector. For the hierarchical response vectors that are valid for the master control label, the response is retained to obtain the master control hierarchical response vector; for the hierarchical response vectors that are valid for the combined label, the hierarchical response vectors corresponding to adjacent risk factor fields in the same combined risk factor are extracted and then spliced ​​dimension by dimension according to the causal direction to obtain the combined hierarchical response vector. Perform scaling reduction on hierarchical response vectors where both master and combined tags are invalid to obtain supplementary hierarchical response vectors; The master hierarchical response vector, combined hierarchical response vector, and supplementary hierarchical response vector are rearranged according to field order to generate a hierarchical evidence response sequence.

[0013] Optionally, S7 specifically includes: S71. The channel assignment of each risk level evidence value in the risk evidence vector of the exploration object is determined, and the low-risk evidence value, medium-risk evidence value, high-risk evidence value and extremely high-risk evidence value are respectively assigned to the corresponding risk level channel. S72. Normalize the evidence values ​​in each risk level channel to obtain the risk attribution value corresponding to each risk level, and determine the risk level with the largest risk attribution value as the risk level of the exploration object. S73. Accumulate all risk level evidence values ​​in the risk evidence vector of the exploration object to obtain the total amount of risk evidence, and determine the risk uncertainty based on the total amount of risk evidence; S74. Match the main control risk factors and combined risk factors in the risk interpretation results with the risk factor stratification results corresponding to the exploration object, extract the main control risk factors that fall into the dominant stratification interval and the combined risk factors that generate combined enhancement, and configure risk source markers. S75. Backfill the risk level, risk uncertainty and risk source markers into the spatial unit of the target exploration area to obtain the geological exploration risk zoning results. Then, prioritize each exploration object in order of risk level from high to low, risk uncertainty from low to high, and the number of combined risk factors from many to few to obtain the exploration priority results.

[0014] The beneficial effects of this invention are: 1. This invention sets up an event projection correction function in the improved optimal parameter geographic detector algorithm, embeds the location mapping relationship of historical geological risk events into the candidate stratification process, and performs score reduction on candidate stratification results that are inconsistent with the distribution of historical geological risk events. This makes the final risk factor stratification results not only have spatial differentiation interpretation ability, but also reflect the actual clustering state of historical geological risk events, reducing the risk of misjudgment caused by relying solely on statistical differentiation.

[0015] 2. This invention introduces a geological genealogy through a target hierarchical function, imposes causal restrictions on risk factor combinations, masks risk factor combinations that do not have a geological genealogical relationship, and retains risk factor combinations that have a genealogical relationship. This makes the identification process of risk factors based on geological genesis, and improves the traceability and rationality of risk source explanation.

[0016] 3. In the improved evidence neural network model, the present invention sets up a hierarchical evidence embedding layer, which embeds the risk factor hierarchical results, risk factor explanatory strength, master control label and combination label into the evidence generation process. This makes the risk level no longer directly mapped from the original risk factor values, but formed by the hierarchical position, explanatory strength and causal combination relationship to form a risk evidence vector, thereby simultaneously outputting the risk level, risk uncertainty and risk source. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a geological exploration risk assessment method based on big data analysis proposed in this invention; Figure 2 This is a flowchart of the improved optimal parameter geographic detector algorithm for a geological exploration risk assessment method based on big data analysis proposed in this invention. Figure 3This is a structural diagram of an improved evidence neural network model containing a hierarchical evidence embedding layer, which is proposed in this invention as a geological exploration risk assessment method based on big data analysis. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A geological exploration risk assessment method based on big data analysis includes the following steps: S1. Aggregate multi-source big data on geological exploration in the target exploration area, and organize it into a dataset of exploration objects and a mapping relationship between the locations of historical geological risk events according to the exploration objects; S2. Configure risk factor fields based on the exploration object dataset, construct a geological risk factor matrix, and generate a geological genetic map based on the geological genetic relationship between the risk factor fields. S3. Input the risk factors in the geological risk factor matrix into the improved optimal parameter geographic detector algorithm, construct the event projection correction function, embed the location mapping relationship of historical geological risk events into the candidate stratification process, construct the target stratification function, introduce the geological gene map to impose causal constraints on the combination of risk factors, and obtain the risk factor stratification results and the risk factor interpretation intensity. S4. Analyze the risk control relationship of the risk factor stratification results, identify the risk factors whose explanatory strength meets the control screening conditions as the main control risk factors, and identify the combination of risk factors along the genetic inheritance path in the geological genetic map to form the risk interpretation results. S5. Rearrange the risk factor codes based on the risk factor stratification results and risk interpretation results to form a risk assessment feature sequence; S6. Input the risk assessment feature sequence into the improved evidence neural network model containing a hierarchical evidence embedding layer to obtain the risk evidence vector of the exploration object. S7. Determine the risk level and risk uncertainty based on the risk evidence vector of the exploration object, determine the source of risk in combination with the risk interpretation results, and backfill to form the geological exploration risk zoning results and exploration priority results.

[0020] In this embodiment, S1 specifically includes: S11. Project the boundary of the target exploration area, the layout of exploration operations, and the existing exploration control positions onto a unified spatial reference, delineate the set of exploration objects, and assign object numbers, object boundaries, and object type markers to each exploration object; S12. Perform source identification, spatial attribution identification, and time period attribution identification on multi-source big data of geological exploration, and link the data that falls within the object boundary and the data that has an operational control relationship with the exploration object to the corresponding exploration object respectively; S13. Perform coverage segmentation and object attribution assignment on data spanning multiple exploration objects, and encapsulate the data that has completed the attribution assignment under the same exploration object within the object to obtain the exploration object dataset; S14. Determine the location and impact range of historical geological risk events, and link them to the corresponding exploration objects to form a location mapping relationship of historical geological risk events.

[0021] In the specific implementation process, the exploration object serves as a spatial unit within the target exploration area, carrying multi-source big data of geological exploration. First, the boundary of the target exploration area, the layout of exploration operations, and existing exploration control locations are unified under the same spatial benchmark. Then, the set of exploration objects is delineated according to the exploration grid, profile segments, borehole control range, and anomaly control range, and each exploration object is assigned an object number and object boundary. The object number participates in the construction of the row index of the geological risk factor matrix, and the object boundary participates in the spatial attribution determination of multi-source data.

[0022] Geological exploration multi-source big data is linked to corresponding exploration objects according to spatial inclusion relationships and operational control relationships. Data items located entirely within the boundaries of a single object are directly assigned to that exploration object; data items with control relationships with boreholes, profile lines, sampling points, or engineering operation areas are assigned to the exploration object according to the corresponding operational control object. This processing method avoids the problem of inaccurate attribution in borehole logging, profile interpretation, and engineering records caused by relying solely on spatial clipping.

[0023] For data items spanning multiple exploration objects, coverage area segmentation is performed first, followed by object attribution. Areal data is segmented into multiple object fragments based on the overlapping area with the boundaries of each exploration object; linear data is segmented into multiple line segments based on the intersection with the boundaries of each exploration object; raster data is assigned to corresponding exploration objects based on cell center location or cell coverage ratio; point data is assigned to corresponding exploration objects based on point location relationships. After segmentation, each object fragment retains its original data number and is assigned an object number and coverage ratio, enabling the contribution range of the same original data item across multiple exploration objects to be distinguished.

[0024] Historical geological risk events are assigned based on their location and impact range. Historical geological risk events with a clearly defined location are linked to corresponding exploration objects according to their locational relationships. Historical geological risk events with an impact range are mapped by overlaying the impact range with the boundary of the exploration object, and the exploration objects covered by the impact range are marked as associated objects. This establishes a locational mapping relationship for historical geological risk events, providing a locational basis for the event projection correction function to project historical geological risk events to candidate risk layers.

[0025] In this embodiment, S2 specifically includes: S21. Classify the data items in the exploration object dataset according to the causes of geological structure, stratigraphy, abnormal response, hydrological environment, engineering disturbance and historical events, and configure the risk factor field set. S22. Perform field repositioning on the data items in each exploration object, convert data items under the same causal category into corresponding risk factor field values, and configure data validity markers for missing fields. S23. Using the exploration object number as the matrix row index and the risk factor field as the matrix column index, the risk factor field values ​​and data validity markers are assembled in the same position to obtain the geological risk factor matrix. S24. The risk factor fields are processed into nodes according to their causal categories, and the geological causal relationship between the risk factor fields is converted into directed relationship edges to obtain the initial geological causal relationship diagram. S25. Verify the endpoints, directions, and duplicates of the directed edges in the initial geological genetic relationship diagram, retaining the directed edges with complete endpoints and consistent directions to obtain the geological genetic map.

[0026] In the specific implementation process, the risk factor field set is used to convert exploration data from different sources and in different forms into a unified risk expression field. First, data items in the exploration object dataset are categorized into fields based on their causal categories, such as geological structure, stratigraphic lithology, anomaly response, hydrological environment, engineering disturbance, and historical events. For example, fault density, fold development degree, and tectonic intersection location are categorized into geological structure fields; lithological weakness, fracture zone thickness, and weathering degree are categorized into stratigraphic lithology fields; and geophysical anomaly intensity, geochemical anomaly concentration, and remote sensing anomaly patches are categorized into anomaly response fields. After field categorization, data items from different sources but with the same causal category within the same exploration object are converted into corresponding risk factor field values, giving each exploration object a unified field structure.

[0027] Field alignment is used to resolve inconsistencies in data sources and data formats within the same exploration object. For numerical data, the data values ​​corresponding to the same field within the object are statistically summarized to obtain the field value; for ordinal data, the ordinal results are converted into coded values ​​that can participate in matrix assembly; for spatial extent data, it is converted into field values ​​based on the coverage ratio or influence range percentage within the object. Fields without valid data are configured with data validity markers to distinguish between "low risk factor value" and "data missing for this field". After field alignment is completed, the risk factor field values ​​and data validity markers are assembled in the same position using the exploration object number as the matrix row index and the risk factor field as the matrix column index to obtain the geological risk factor matrix.

[0028] Risk factor field nodeization transforms the field relationships in the matrix into node relationships in a graph structure. Each risk factor field is configured as a causal graph node, which retains the field name, causal category, and field source marker. Geological causal relationships describe the direction of geological interaction transmission between risk factors; for example, geological structure fields point to stratigraphy / lithology fields or hydrological / environmental fields, hydrological / environmental fields point to historical event fields, and engineering disturbance fields point to historical event fields. These causal relationships are converted into directed edges, with the starting point of each edge corresponding to an upstream causal field and the ending point corresponding to a downstream response field, thus obtaining the initial geological causal relationship graph.

[0029] Endpoint verification checks whether both the start and end fields of each directed edge exist in the risk factor field set; if either the start or end field is missing, the directed edge is not included in the geological genetic map. Direction verification checks whether the direction of the directed edge conforms to the geological genetic sequence, for example, from tectonic control to hydrological response, or from hydrological response to risk event, rather than a reverse connection. Duplicate merging handles situations where multiple data sources provide the same start and end fields; it retains one directed edge and merges duplicate sources into an edge source tag. After completing endpoint verification, direction verification, and duplicate merging, directed edges with complete endpoints, consistent directions, and no duplicate conflicts are retained to obtain the geological genetic map.

[0030] In this embodiment, S3 specifically includes: S31. In the improved optimal parameter geographic detector algorithm, each risk factor field in the geological risk factor matrix is ​​expanded into a field value sequence, and candidate discrete segmentation is performed on the field value sequence to obtain a candidate stratification result set. S32. Perform interval assignment calibration on each candidate stratification result in the candidate stratification result set, label the stratification interval in each candidate stratification result as a candidate risk layer, and assign each exploration object to the corresponding candidate risk layer according to the corresponding risk factor value, and calculate the initial stratification score value of the candidate stratification result according to the factor detection rules of the geographic detector. S33. Project the location mapping relationship of historical geological risk events onto the candidate risk layer, count the density value of historical geological risk events in each candidate risk layer, and calculate the event density difference value between adjacent candidate risk layers. S34. Construct an event projection correction function. The event projection correction function takes the initial stratification score value, the historical geological risk event density value and the event density difference value as input. It performs score retention on the candidate stratification results where the historical geological risk events are concentrated in the high-risk layer, and performs score reduction on the candidate stratification results where the distribution of historical geological risk events is inconsistent with the candidate risk level. It outputs the event-corrected stratification score value. S35. Convert the geological genetic map into a risk factor causal adjacency table, configure effective causal adjacency values ​​for risk factor combinations with geological genetic inheritance relationships, and configure invalid causal adjacency values ​​for risk factor combinations without geological genetic inheritance relationships. S36. Construct a target stratification function. The target stratification function takes the event-corrected stratification score and the risk factor causal adjacency value as input. It performs masking on the candidate stratification results corresponding to the invalid causal adjacency values, retains the candidate stratification results corresponding to the valid causal adjacency values, and selects the target stratification result from the retained candidate stratification results. S37. The stratification boundary, stratification number, and score value corresponding to the target stratification result are determined as the risk factor stratification result and the risk factor explanatory strength.

[0031] In the specific implementation process, the stratification results and explanatory strength of each risk factor are determined by improving the optimal parameter geographic detector algorithm. The improved optimal parameter geographic detector algorithm does not simply perform discretization processing on the risk factor fields, but adds an event projection correction function and a target stratification function during the candidate stratification search process, so that the candidate stratification results are simultaneously constrained by the spatial differentiation of risk, the distribution of historical geological risk events, and the geological genetic inheritance relationship.

[0032] Each risk factor field in the geological risk factor matrix is ​​expanded into a sequence of field values, arranged according to the exploration object number. The values ​​of the same risk factor field across different exploration objects are recorded. Candidate discrete segmentation is performed on the field value sequences, including equidistant segmentation, quantile segmentation, natural discontinuity segmentation, and cluster segmentation. Multiple candidate stratification results are generated under different numbers of strata. Each candidate stratification result includes the stratification boundary, stratification number, and the stratification affiliation corresponding to each exploration object. Multiple candidate stratification results together constitute a candidate stratification result set.

[0033] For each candidate stratification result in the candidate stratification result set, interval assignment is performed, designating each stratification interval in the candidate stratification result as a candidate risk layer, and each exploration object is assigned to the corresponding candidate risk layer according to the value of the corresponding risk factor. The risk direction of the candidate risk layer is determined by the risk direction labeling of the risk factor field; for fields such as fracture density, anomaly response intensity, hydrological disturbance intensity, and engineering disturbance intensity, where higher values ​​indicate higher risk, high-value stratification intervals are designated as high-risk layers; for fields such as rock mass integrity and stable rock layer thickness, where lower values ​​indicate higher risk, low-value stratification intervals are designated as high-risk layers. Through this processing, the candidate risk layers can reflect the true risk direction of different risk factor fields.

[0034] The initial stratification score is calculated according to the geographic detector factor detection rules. First, a risk response value is assigned to each exploration object. This risk response value is derived by aggregating the anomaly response field, hydrological environment field, engineering disturbance field, and historical event field, and is used to represent the risk status of the exploration object. For any candidate stratification result, the number of exploration objects, the mean risk response, and the risk response dispersion within each candidate risk layer are statistically analyzed, and the overall risk response dispersion of all exploration objects is also statistically analyzed. If the risk response dispersion within a candidate risk layer is low, and the mean risk response differs significantly between different candidate risk layers, it indicates that the candidate stratification result can effectively distinguish exploration objects with different risk statuses, corresponding to a higher initial stratification score. Conversely, if the risk response dispersion within a candidate risk layer is high, or the mean risk response is similar between different candidate risk layers, it indicates that the candidate stratification result has a weak explanatory power for spatial differences in risk, corresponding to a lower initial stratification score.

[0035] The event projection correction function serves as a scoring correction subroutine within the improved optimal parameter geographic detector algorithm. Its inputs include initial stratified score values, historical geological risk event density values, and event density difference values; the output is the corrected stratified score value. The function projects the location mapping relationships of historical geological risk events onto candidate risk layers, ensuring that each historical geological risk event falls into its corresponding candidate risk layer according to the associated exploration object. Subsequently, the number of historical geological risk events within each candidate risk layer is counted, and this, combined with the number of exploration objects within each candidate risk layer, yields the historical geological risk event density value. Finally, the historical geological risk event densities between adjacent candidate risk layers are compared to obtain the event density difference value.

[0036] The event projection correction function adjusts the initial stratification score based on the distribution of historical geological risk events in candidate risk layers. When historical geological risk events are mainly concentrated in high-risk layers, and there is a significant difference in event density between high-risk layers and adjacent low-risk layers, the event projection correction function retains the score for the candidate stratification result, allowing the initial stratification score to continue as a valid stratification score in the target stratification selection. When the distribution of historical geological risk events is inconsistent with the candidate risk level, for example, the event density in low-risk layers is higher than in high-risk layers, or the difference in event density between different candidate risk layers is insufficient to distinguish risk levels, the event projection correction function reduces the score of the candidate stratification result. The score-reduced result is output as the event-corrected stratification score value, used to reduce the probability of candidate stratification results that are inconsistent with the distribution of historical geological risk events being selected.

[0037] The target stratification function serves as the target selection subroutine within the improved optimal parameter geographic detector algorithm. Its input values ​​include event-corrected stratification scores and risk factor causal adjacency values, with the output being the target stratification result. The geological genetic map is converted into a risk factor causal adjacency table, where risk factor fields are used as rows and columns. When a geological genetic relationship exists between two risk factor fields, the corresponding field combination is configured with valid causal adjacency values; when no geological genetic relationship exists between two risk factor fields, the corresponding field combination is configured with invalid causal adjacency values.

[0038] The target stratification function masks candidate stratification results corresponding to invalid causal adjacency values, preventing risk factor combinations without geological genetic support from participating in the target stratification result selection. It retains candidate stratification results corresponding to valid causal adjacency values, allowing risk factor combinations with tectonic control, lithological response, hydrological conduction, engineering disturbance, or historical event inheritance relationships to participate in the target stratification result selection. After completing the causal adjacency screening, the target stratification function compares the event-corrected stratification scores among the retained candidate stratification results and selects those whose scores meet the target stratification criteria as the target stratification results. The target stratification criteria can be determined by the score ranking of candidate stratification results under the same risk factor field, or by the difference between the score and the average score of the candidate stratification results. The specific values ​​are configured based on the data scale of the target exploration area and the number of historical events during implementation.

[0039] The stratification boundaries and stratification numbers corresponding to the target stratification results are determined as the risk factor stratification results, and the score values ​​corresponding to the target stratification results are determined as the explanatory strength of the risk factors. The resulting risk factor stratification results not only reflect the explanatory power of risk factor values ​​for the spatial differences in risk of the exploration object, but also, after historical geological risk event projection correction and geological genetic map causal screening, can reduce the number of candidate stratification results without historical event support or causal relationship support entering the risk interpretation process, providing reliable input for risk control relationship analysis and improvement of evidence neural network models.

[0040] In this embodiment, S4 specifically includes: S41. Assign intervals to each risk factor in the risk factor stratification results according to the stratification number, and assign the explanatory strength of the risk factor to the corresponding stratification interval to obtain the single-factor risk explanation record. S42. Sort the risk factors in the single-factor risk interpretation record according to their explanatory strength. Mark the risk factors whose explanatory strength meets the main control screening criteria and whose historical geological risk event density in the corresponding stratified interval meets the event concentration criteria as the main control risk factors. S43. Locate the causal path of the main control risk factors in the geological causal map, connect the main control risk factors with geological causal relationship to obtain the candidate risk factor combination. S44. Cross-over the risk factor stratification intervals in the candidate risk factor combination, classify the exploration objects falling into the same overlapping interval into the same combination risk unit, and calculate the combination explanation strength of the combination risk unit. S45. Compare the combined explanatory strength of the combined risk unit with the corresponding single-factor risk explanatory record, configure a combined risk-causing label for candidate risk-causing factors whose explanatory strength is enhanced after combination, and determine the combined risk-causing factors. S46. Encapsulate the main control risk factors, main control stratification intervals, combined risk factors, combined risk units, and geological origination paths into risk interpretation results.

[0041] In the specific implementation process, the stratification number, stratification boundary, explanatory strength of the risk factor, and density of historical geological risk events within the stratification interval are linked to obtain a single-factor risk interpretation record. The controlling factor screening criteria are determined by the explanatory strength of the risk factor. Specifically, all risk factors can be sorted from high to low according to their explanatory strength, and the top 30% of risk factors are selected as candidate controlling risk factors. When the number of risk factors is small, risk factors with an explanatory strength higher than the average explanatory strength of all risk factors can also be selected as candidate controlling risk factors. The event concentration criteria are determined by the distribution of historical geological risk events within the stratification interval. Specifically, the density of historical geological risk events in the high-risk stratification interval is higher than the average event density of the target exploration area and higher than the density of historical geological risk events in adjacent low-risk stratification intervals. Risk factors that meet both the controlling factor screening criteria and the event concentration criteria are marked as controlling risk factors.

[0042] The controlling risk factors are mapped to nodes in the genetic map of the geological genetic atlas, and the genetic inheritance paths between the controlling risk factors are located along the directed relation edges. Controlling risk factors with genetic inheritance paths are connected by the paths to form candidate risk factor combinations. When candidate risk factor combinations are cross-over overlaid, the stratification numbers of the same exploration object under each controlling risk factor are combined into a combined stratification state. Exploration objects with the same combined stratification state are assigned to the same combined risk unit. For example, if an exploration object falls into both the structurally fractured high-risk layer and the hydrologically disturbed high-risk layer, it is assigned to the corresponding high-risk combined unit.

[0043] The combined interpretation strength is used to measure the explanatory power of a combined risk unit for the distribution of historical geological risk events. The processing equipment statistically analyzes the number of exploration objects, the number of historical geological risk events, and the event density within each combined risk unit, and compares the differences in event density between different combined risk units. The more concentrated the event distribution within a combined risk unit and the more significant the differences in event density between combined risk units, the higher the combined interpretation strength. The combined interpretation strength is compared with the risk interpretation records of individual factors within the combination. When the combined interpretation strength is higher than the risk factor interpretation strength of individual factors within the combination, or when the increase in the relative highest single-factor interpretation strength reaches the configured enhancement judgment range, a combined risk-causing marker is configured for the candidate risk-causing factor combination. The controlling risk factor, controlling stratification interval, combined risk-causing factor, combined risk unit, and geological genetic inheritance path are then encapsulated as the risk interpretation result.

[0044] In this embodiment, S5 specifically includes: S51. Match the stratification number in the risk factor stratification result with the main control risk factor and combined risk-causing factor in the risk interpretation result, and configure the main control label, combined label and stratification label for the risk factor field in the geological risk factor matrix; S52. Arrange the risk factor fields with the configured master control tags first, and determine the field order from high to low according to the explanatory strength of the risk factors; S53. Map the risk factor fields of the configured combination markers to the directed relation edges in the geological genealogy, extract the starting risk factor fields and the ending risk factor fields corresponding to the same combination of risk factors, and arrange them adjacently in the order of starting risk factor fields first and ending risk factor fields last to form a combination field segment. S54. Risk factor fields without configured master control and combination labels are retained in the supplementary field section and listed after the master control and combination field sections; S55. Encapsulate the risk factor values, stratification numbers, risk factor interpretation strengths, main control markers, and combination markers corresponding to each exploration object in the same position according to the rearranged field order to form a risk assessment feature sequence.

[0045] In the specific implementation process, each risk factor field is first matched with its corresponding stratification number, risk factor explanatory strength, master risk factor record, and combined risk-causing factor record, and a master control tag, combined tag, and stratification tag are configured for the risk factor field. When the same risk factor field has both a master control tag and a combined tag, the risk factor field is first retained in the master control field segment, while the combined field segment retains the combined relationship index between the risk factor field and the adjacent risk-causing field, to avoid the same field participating in the embedding calculation repeatedly.

[0046] Field rearrangement follows the order of "main control field segment, combined field segment, and supplementary field segment". In the main control field segment, risk factor fields are arranged from high to low explanatory power, placing risk factors with higher explanatory power at the beginning of the risk assessment feature sequence. In the combined field segment, risk factor fields with combined labels are mapped to directed relation edges in the geological genetic map and arranged adjacently with the starting risk factor field first and the ending risk factor field last, ensuring continuity of fields corresponding to the same combined risk factors in the sequence. The supplementary field segment is used to retain risk factor fields that have not yet entered the risk interpretation results and is arranged after the main control field segment and the combined field segment.

[0047] Co-location encapsulation refers to placing the risk factor value, stratification number, risk factor explanatory strength, main control label, and combined label into the same field feature unit for each exploration object and each rearranged risk factor field. Multiple field feature units, arranged in the rearranged field order, constitute the risk assessment feature sequence corresponding to that exploration object. This process enables the stratified evidence embedding layer to simultaneously obtain the original risk factor value, stratification position, explanatory strength, and risk interpretation label, avoiding the neural network's evidence mapping based solely on the original risk factor value.

[0048] In this embodiment, the improved evidence neural network model specifically includes a risk feature embedding layer, a hierarchical evidence embedding layer, a risk latent representation mapping layer, and an evidence output layer: The risk feature embedding layer divides the risk assessment feature sequence into field feature units according to the order of risk factor fields. Each field feature unit includes risk factor code, hierarchical number, risk factor explanatory strength, master control label and combination label. The risk factor codes in the field feature units are expanded into one-dimensional factor code vectors, and the stratified number, main control label, and combined label are converted into stratified label vectors, main control label vectors, and combined label vectors, respectively. The one-dimensional factor encoding vector, hierarchical label vector, master control label vector and combined label vector are concatenated in a fixed order within the fields to obtain the field embedding vector, and the field embedding vectors are arranged in the order of the risk factor fields to form a field embedding sequence. The hierarchical evidence embedding layer performs field matching and hierarchical position matching between the field embedding vectors in the field embedding sequence and the risk factor hierarchical results. It then performs hierarchical anchoring and evidence response transformation on the matched field embedding vectors to generate a hierarchical evidence response sequence. The risk implicit representation mapping layer divides the hierarchical evidence response sequence into the main control evidence segment, the combined evidence segment, and the supplementary evidence segment according to the main control label and the combined label. The hierarchical evidence response vector in each evidence segment is accumulated dimension by dimension to obtain the main control evidence vector, the combined evidence vector, and the supplementary evidence vector. The main evidence vector, combined evidence vector, and supplementary evidence vector are concatenated in sequence and then processed by linear mapping to obtain the risk hidden representation of the exploration object; The evidence output layer maps the implicit risk representation of the exploration object to the evidence channels corresponding to low risk, medium risk, high risk and extremely high risk, and arranges the output values ​​of each evidence channel in order of risk level to obtain the risk evidence vector of the exploration object.

[0049] In practical implementation, ordinary evidence neural networks typically encode risk factor values ​​directly and feed them into the hidden layer, with the evidence output layer then generating evidence values ​​for each risk level. This approach relies primarily on network weights learning the correspondence between risk factors and risk levels. The risk factor stratification results, risk factor explanatory strength, controlling risk factors, and combined risk-causing factors lack clear structural constraints within the network, making the risk evidence generation process prone to black-box mapping. The improved evidence neural network model adds a stratified evidence embedding layer between the risk feature embedding layer and the risk hidden representation mapping layer of the ordinary evidence neural network. This allows the risk factor stratification results and risk factor explanatory strength obtained from the improved optimal parameter geographic detector algorithm to enter the network's internal evidence response process, thereby transforming the "statistical stratified explanatory results" into "network evidence generation constraints."

[0050] The improved evidence neural network model includes a risk feature embedding layer, a hierarchical evidence embedding layer, a risk latent representation mapping layer, and an evidence output layer. The risk feature embedding layer converts the risk assessment feature sequence into a field embedding sequence; the hierarchical evidence embedding layer anchors and aligns the field embedding vectors with the risk factor hierarchical results, and generates hierarchical evidence responses based on the explanatory strength of the risk factors; the risk latent representation mapping layer divides the hierarchical evidence response sequence into main control evidence segments, combined evidence segments, and supplementary evidence segments, and merges each evidence segment into the risk latent representation of the exploration object; the evidence output layer maps the risk latent representation of the exploration object into evidence values ​​corresponding to low risk, medium risk, high risk, and extremely high risk.

[0051] The risk feature embedding layer comprises field segmentation units, factor encoding units, label encoding units, and field arrangement units. The field segmentation unit segments the risk assessment feature sequence according to the order of risk factor fields, encapsulating the data corresponding to each risk factor field into a field feature unit. This field feature unit includes the risk factor code, stratification number, risk factor explanatory strength, master control label, and combination label. The factor encoding unit performs vectorization processing on the risk factor codes, normalizing numerical risk factors by numerical range and expanding them into one-dimensional factor encoding vectors, and converting categorical risk factors into one-dimensional factor encoding vectors by category index. The label encoding unit converts the stratification number, master control label, and combination label into label vectors that can participate in vector operations. The stratification number is converted into a stratification label vector, the master control label into a master control label vector, and the combination label into a combination label vector. The field arrangement unit concatenates the risk factor encoding vector, stratification label vector, risk factor explanatory strength vector, master control label vector, and combination label vector in that order to obtain the field embedding vector; these field embedding vectors are then rearranged according to the field rearrangement order in the risk assessment feature sequence to form the field embedding sequence.

[0052] The hierarchical evidence embedding layer is the core improvement layer of the improved evidence neural network model compared to the ordinary evidence neural network model. This layer includes a field matching unit, a hierarchical anchoring unit, an explanatory strength correction unit, and an evidence response triage unit. The field matching unit performs field matching and hierarchical position matching between the field embedding vector in the field embedding sequence and the risk factor hierarchical results to determine the risk factor field and hierarchical number corresponding to the field embedding vector. The hierarchical anchoring unit extracts the hierarchical anchoring vector corresponding to the risk factor field and hierarchical number from the hierarchical anchoring table, and adds the field embedding vector to the hierarchical anchoring vector dimension-wise to obtain the hierarchical alignment vector. The explanatory strength correction unit expands the risk factor explanatory strength according to the dimensions of the hierarchical alignment vector into an explanatory strength correction vector, and multiplies the hierarchical alignment vector to the explanatory strength correction vector dimension-wise to obtain the hierarchical response vector. The evidence response diversion unit diverts the hierarchical response vectors according to the master control marker and the combined marker. For hierarchical response vectors with valid master control markers, the response is retained. For hierarchical response vectors with valid combined markers, they are concatenated dimension by dimension with the hierarchical response vectors corresponding to adjacent risk factor fields according to the causal direction. For hierarchical response vectors with invalid master control markers and combined markers, the proportional reduction is performed, and finally, a hierarchical evidence response sequence is generated.

[0053] The effect of the hierarchical evidence embedding layer lies in changing the way ordinary evidence neural networks freely map risk factors. In ordinary evidence neural networks, risk factors typically enter the hidden layer as parallel features, making it difficult for the network to distinguish between high-explanatory-strength risk factors, dominant risk factors, and supplementary risk factors. Through the hierarchical anchoring unit, the field embedding vector is forced to align to the hierarchical number determined by the improved optimal parameter geographic detector algorithm; through the explanatory-strength correction unit, risk factors with higher explanatory-strength are amplified in the evidence response, while the response of risk factors with lower explanatory-strength is relatively weakened; through the evidence response diversion unit, dominant risk factors, combined risk factors, and supplementary risk factors enter different response paths. Thus, the risk evidence vector is not simply obtained by mapping the original risk factor values, but is jointly determined by the risk factor values, hierarchical position, explanatory strength, and causal combination relationship.

[0054] In practical implementation, the risk implicit representation mapping layer includes evidence segment division units, intra-segment merging units, and implicit representation mapping units. The evidence segment division unit divides the hierarchical evidence response sequence into main control evidence segments, combined evidence segments, and supplementary evidence segments based on the main control marker and combination marker. The main control evidence segment contains the hierarchical evidence response vectors corresponding to the main control risk factors, the combined evidence segment contains the combined hierarchical response vectors corresponding to the combined risk-causing factors, and the supplementary evidence segment contains the supplementary hierarchical response vectors that did not enter the risk interpretation results. The intra-segment merging unit performs dimension-wise accumulation of the hierarchical evidence response vectors in each evidence segment to obtain the main control evidence vector, combined evidence vector, and supplementary evidence vector. The implicit representation mapping unit concatenates the main control evidence vector, combined evidence vector, and supplementary evidence vector in sequence and performs linear mapping processing to obtain the implicit representation of the exploration object's risk. The function of this layer is to preserve the structural differences between the main control risk, combined risk, and supplementary risk, and to avoid indiscriminate mixing of evidence responses from different sources in the same implicit space.

[0055] The evidence output layer comprises risk level evidence channels and evidence arrangement units. Risk level evidence channels correspond to low risk, medium risk, high risk, and extremely high risk, respectively. Each channel receives the implicit risk representation of the exploration object and outputs the corresponding evidence value for its risk level. The evidence arrangement unit arranges the output values ​​of each evidence channel in the order of low risk, medium risk, high risk, and extremely high risk, resulting in a risk evidence vector for the exploration object. The evidence output layer does not directly provide a single risk level; instead, it first outputs the evidence values ​​corresponding to each risk level, ensuring that the risk level and risk uncertainty can be calculated based on the evidence values. When the evidence values ​​for each risk level differ significantly and the total amount of evidence is high, the risk level judgment is more stable; when the evidence values ​​are scattered or the total amount of evidence is low, the risk uncertainty increases, which is beneficial for identifying objects requiring further exploration.

[0056] In this embodiment, the field embedding vectors in the field embedding sequence are matched with the risk factor stratification results for field matching and stratification position matching. The matched field embedding vectors are then subjected to stratification anchoring and evidence response transformation to generate a stratified evidence response sequence, specifically including: The risk factor stratification results are split into field stratification records according to the risk factor field. Each field stratification record includes the risk factor field, stratification number, stratification boundary, and risk factor explanatory strength. A stratification anchor vector is configured for each stratification number, and a stratification anchor table is established. For the target field embedding vector in the field embedding sequence, extract the corresponding risk factor field and stratification number, match the risk factor field and stratification number with the stratification anchoring table, and locate the target stratification anchoring vector. The target field embedding vector and the target hierarchical anchoring vector are added dimension by dimension to generate a hierarchical alignment vector; The explanatory strength of the risk factors corresponding to the target field embedding vector is expanded into an explanatory strength correction vector according to the dimensions of the hierarchical alignment vector. The hierarchical alignment vector and the explanatory strength correction vector are then multiplied dimension by dimension to generate a hierarchical response vector. For the hierarchical response vectors that are valid for the master control label, the response is retained to obtain the master control hierarchical response vector; for the hierarchical response vectors that are valid for the combined label, the hierarchical response vectors corresponding to adjacent risk factor fields in the same combined risk factor are extracted and then spliced ​​dimension by dimension according to the causal direction to obtain the combined hierarchical response vector. Perform scaling reduction on hierarchical response vectors where both master and combined tags are invalid to obtain supplementary hierarchical response vectors; The master hierarchical response vector, combined hierarchical response vector, and supplementary hierarchical response vector are rearranged according to field order to generate a hierarchical evidence response sequence.

[0057] In practical implementation, the hierarchical evidence embedding layer is used to embed the risk factor stratification results obtained from the improved optimal parameter geographic detector algorithm into the evidence generation process. The hierarchical anchor vector is configured according to the risk factor field and stratification number; different stratification numbers under the same risk factor field correspond to different hierarchical anchor vectors, used to distinguish the positions of low-risk, medium-risk, high-risk, and extremely high-risk strata in the evidence space. After the target field embedding vector and the target hierarchical anchor vector are added dimension-wise, the field embedding vector obtains the corresponding stratification position constraint, generating a hierarchical alignment vector. The explanatory strength of the risk factors is first normalized to a uniform numerical range, then expanded and replicated along the dimensions of the hierarchical alignment vector to form an explanatory strength correction vector, which is then multiplied dimension-wise with the hierarchical alignment vector, so that risk factors with higher explanatory strength receive a stronger response in the hierarchical response vector. For stratified response vectors with valid master control markers, they are directly retained as master control stratified response vectors. For stratified response vectors with valid combined markers, they are concatenated dimension-by-dimensional with the stratified response vectors corresponding to adjacent risk factor fields according to the genetic continuity direction in the geological genetic map to obtain combined stratified response vectors. For stratified response vectors with invalid master control markers and combined markers, they are proportionally reduced to supplementary stratified response vectors. The reduction ratio can be determined based on the number of supplementary risk factors and the explanatory strength distribution of the master control risk factors, for example, a reduction coefficient in the range of 0.3 to 0.6. Through the above processing, the stratified evidence response sequence can simultaneously retain the stratification position, explanatory strength, and risk explanatory role of risk factors, avoiding excessive interference from supplementary risk factors on the evidence responses of master control risk factors and combined risk factors.

[0058] In this embodiment, S7 specifically includes: S71. The channel assignment of each risk level evidence value in the risk evidence vector of the exploration object is determined, and the low-risk evidence value, medium-risk evidence value, high-risk evidence value and extremely high-risk evidence value are respectively assigned to the corresponding risk level channel. S72. Normalize the evidence values ​​in each risk level channel to obtain the risk attribution value corresponding to each risk level, and determine the risk level with the largest risk attribution value as the risk level of the exploration object. S73. Accumulate all risk level evidence values ​​in the risk evidence vector of the exploration object to obtain the total amount of risk evidence, and determine the risk uncertainty based on the total amount of risk evidence; S74. Match the main control risk factors and combined risk factors in the risk interpretation results with the risk factor stratification results corresponding to the exploration object, extract the main control risk factors that fall into the dominant stratification interval and the combined risk factors that generate combined enhancement, and configure risk source markers. S75. Backfill the risk level, risk uncertainty and risk source markers into the spatial unit of the target exploration area to obtain the geological exploration risk zoning results. Then, prioritize each exploration object in order of risk level from high to low, risk uncertainty from low to high, and the number of combined risk factors from many to few to obtain the exploration priority results.

[0059] In practice, the risk evidence vector for the exploration object includes low-risk, medium-risk, high-risk, and extremely high-risk evidence values. The processing equipment first sums the evidence values ​​for the four risk levels to obtain a risk evidence sum. Then, it divides each risk level evidence value by the risk evidence sum to obtain the risk attribution value for that risk level. The risk level with the highest risk attribution value is determined as the risk level of the exploration object. When the risk evidence sum is zero or lower than the minimum evidence determination value, the risk attribution values ​​for each risk level are configured to an equilibrium state, and the exploration object is marked as a high-uncertainty object.

[0060] Risk uncertainty is determined inversely by the total amount of risk evidence. A higher total amount of risk evidence indicates stronger evidence support for the risk level of the exploration object by the improved evidence neural network model, resulting in lower risk uncertainty. Conversely, a lower total amount of risk evidence indicates insufficient or scattered evidence for each risk level, resulting in higher risk uncertainty. Risk uncertainty can be determined in segments. For example, a low uncertainty can be configured when the total amount of risk evidence is above a high evidence threshold, a medium uncertainty when the total amount of risk evidence is between the high and low evidence thresholds, and a high uncertainty when the total amount of risk evidence is below the low evidence threshold. The high and low evidence thresholds can be determined based on the quantiles of the distribution of the total amount of risk evidence in the training samples or historical exploration samples.

[0061] Risk source marking is determined jointly by the risk interpretation results and the risk factor stratification results. The processing equipment matches the main control risk factors with the risk factor stratification results corresponding to the exploration object. If the exploration object falls into the main control stratification interval of the main control risk factor, the corresponding main control risk source marking is configured. It also matches the combined risk factors with the combined risk units of the exploration object. If the exploration object falls into the combined risk unit that generates combined enhancement, a combined risk source marking is configured. After completing the risk source marking, the risk level, risk uncertainty, and risk source marking are backfilled into the spatial units of the corresponding exploration object, forming the geological exploration risk zoning results. Furthermore, exploration priority results are generated according to risk level from high to low, risk uncertainty from low to high, and the number of combined risk factors from many to few.

[0062] Example 1: To verify the feasibility of this invention in practice, it was applied to a geological exploration risk assessment task in a deep exploration area on the periphery of a metal mine. The target exploration area covers approximately 46.8 km², with well-developed fault structures, localized fracture zones, water-rich anomalies, weak lithological interlayers, and historical borehole water inflow records. Traditional risk assessment mainly relies on expert-weighted overlay and single-factor anomaly delineation, which easily leads to the fragmented treatment of geophysical anomalies, structural anomalies, and hydrological anomalies. This results in the underestimation of some high-risk exploration areas and misclassification of areas with overlapping anomalies lacking causal connections as high-risk areas. Especially in areas where fault intersections coincide with hydrological disturbance zones, traditional methods struggle to explain the sources of risk, affecting borehole layout, supplementary exploration sequence, and on-site risk management.

[0063] In this embodiment, the processing equipment aggregates geological mapping data, borehole logging data, geophysical inversion data, geochemical anomaly data, hydrological engineering records, remote sensing interpretation data, and historical geological risk event data for the target exploration area. The target exploration area is divided into 326 exploration objects, including 248 areal exploration objects, 52 linear profile segment exploration objects, and 26 exploration objects within the influence range of point-controlled boreholes. A total of 74 historical geological risk events are recorded, including 21 borehole water inrush events, 18 surrounding rock fracturing and instability events, 9 shallow collapse events, 16 abnormal water-rich exposure events, and 10 construction disturbance anomaly events. Structural lines, geophysical anomaly zones, and remote sensing anomaly patches spanning multiple exploration objects are divided according to object boundaries and participate in object attribution based on coverage ratio, forming a mapping relationship between the exploration object dataset and the location of historical geological risk events.

[0064] In the risk factor construction phase, the processing equipment configures risk factor fields according to six causal categories: geological structure, stratigraphic lithology, anomaly response, hydrological environment, engineering disturbance, and historical events, forming a geological risk factor matrix containing 28 risk factor fields. Geological structure fields include fault proximity, fault density, and structural intersection intensity; stratigraphic lithology fields include rock mass integrity, proportion of weak interlayers, and degree of weathering and fracturing; anomaly response fields include geophysical low resistivity anomaly intensity, magnetic anomaly gradient, and geochemical anomaly concentration; hydrological environment fields include water-bearing index, groundwater disturbance intensity, and aquifer connectivity; engineering disturbance fields include borehole density and existing engineering disturbance intensity; and historical event fields include historical event density and event impact coverage ratio. Subsequently, these risk factor fields are node-based, and a geological causal map is established according to the causal relationship of "structural control—lithological fracturing—hydrological connectivity—risk event triggering."

[0065] In the execution phase of the improved optimal parameter geographic detector algorithm, the processing device expands the field value sequences of 28 risk factor fields and performs equidistant segmentation, quantile segmentation, natural breakpoint segmentation, and clustering segmentation. The number of candidate strata is set to 3, 4, and 5, generating a total of 336 candidate stratification results. After calculating the initial stratification score for each candidate stratification result, the event projection correction function projects 74 historical geological risk events to the corresponding candidate risk layers, and calculates the density value of historical geological risk events within each candidate risk layer and the difference in event density between adjacent candidate risk layers. If the density of historical geological risk events in a high-risk layer is significantly higher than that in a low-risk layer, the score of the corresponding candidate stratification result is retained; if historical geological risk events mainly fall into the low-risk layer, the corresponding score is reduced. The target stratification function further introduces a geological genealogy, masks risk factor combinations without causal relationships, retains risk factor combinations with causal relationships, and selects the target stratification result from the retained results to obtain the risk factor stratification results and risk factor interpretation intensity.

[0066] In the risk control relationship analysis stage, the processing equipment screens the main control risk factors based on the explanatory strength of risk factors and the density of historical geological risk events. In this embodiment, the main control screening criteria are set as risk factors ranking in the top 30% in terms of explanatory strength, and event concentration criteria are set as the density of historical geological risk events in the high-risk stratification interval being higher than the average event density of the target exploration area. Five main control risk factors are obtained through screening: fault proximity, tectonic intersection intensity, geophysical low resistivity anomaly intensity, water-bearing index, and rock mass integrity. Subsequently, combined risk factors are identified along the genetic continuity path in the geological genetic map, among which "fault proximity - water-bearing index", "tectonic intersection intensity - geophysical low resistivity anomaly intensity", and "rock mass integrity - engineering disturbance intensity" are marked as combined risk factors. Through cross-over, the above combined risk factors form 41 combined risk units, of which 17 are high-risk combined units.

[0067] In the risk evidence generation stage, the processing equipment encapsulates the risk factor values, stratification numbers, risk factor explanatory strengths, master control markers, and combination markers in the rearranged field order to form a risk assessment feature sequence corresponding to each exploration object. This risk assessment feature sequence is input to an improved evidence neural network model containing a stratified evidence embedding layer. The stratified evidence embedding layer establishes a stratified anchoring table based on the risk factor stratification results, adds the field embedding vector to the corresponding stratified anchoring vector dimension-by-dimensional, and uses the risk factor explanatory strength to form an explanatory strength correction vector. This vector is then multiplied dimension-by-dimensionally to generate a stratified response vector. The master control risk factor response is retained, the combined risk factor response is concatenated dimension-by-dimensionally according to the causal direction, and the supplementary risk factor response is proportionally reduced using a reduction factor of 0.45, ultimately generating a stratified evidence response sequence. The evidence output layer outputs evidence values ​​for four evidence channels: low risk, medium risk, high risk, and extremely high risk, and determines the risk level and risk uncertainty.

[0068] To verify the evaluation results, this embodiment selected 326 exploration objects as evaluation objects. After expert review, borehole verification, and historical data verification, 128 objects were identified as low-risk, 102 as medium-risk, 66 as high-risk, and 30 as extremely high-risk. The expert weight superposition method, the ordinary optimal parameter geographic detector combined with the traditional evidence neural network method, and the method of this invention were compared. The statistical results are shown in Table 1 below.

[0069] Table 1 Comparison of the Effectiveness of Geological Exploration Risk Assessment

[0070] As shown in Table 1, the method of this invention outperforms the comparative methods in terms of risk level accuracy, recall rate for high and extremely high risks, consistency rate of risk sources, and hit rate of the top 30 priority levels. The expert weighting method, relying mainly on manual weighting and layer overlay, achieved a risk level accuracy of 81.29% and a recall rate of 76.04% for high and extremely high-risk objects, indicating that some high-risk exploration objects were not fully identified. The ordinary optimal parameter geographic detector combined with the traditional evidence neural network method can improve risk identification by utilizing stratified interpretation intensity, increasing the risk level accuracy to 87.42%. However, because it does not embed historical geological risk event projection relationships and geological genetic maps into the candidate stratification selection process, there are still statistically correlated but insufficiently causally supported risk combinations. After adopting the method of this invention, the risk level accuracy reached 93.25%, the high and extremely high-risk recall rate reached 91.67%, and the risk source consistency rate reached 89.88%, indicating that the event projection correction function and the target stratification function can improve the consistency between the risk factor stratification results and the actual risk events and geological genetic relationships.

[0071] Among them, the risk level accuracy rate represents the proportion of all exploration objects whose risk level is correctly determined; the high-risk and extremely high-risk recall rate represents the proportion of exploration objects that are truly high-risk or extremely high-risk and are correctly identified; the risk source consistency rate represents the proportion of risk source markers output by this invention that are consistent with expert review conclusions; and the top 30 priority hit rate represents the proportion of truly high-risk or extremely high-risk objects among the top 30 priority objects in the exploration priority ranking. The top 30 priority hit rate of the method of this invention reaches 90.00%, indicating that high-risk exploration objects can be prioritized, which is beneficial for supplementary exploration, borehole layout, and on-site risk management.

[0072] This embodiment demonstrates that the present invention, by improving the event projection correction function in the optimal parameter geographic detector algorithm, embeds the location mapping relationship of historical geological risk events into the candidate stratification process. This ensures that the candidate stratification results not only possess spatial differentiation interpretation capabilities but also maintain consistency with the distribution of historical geological risk events. Furthermore, by introducing a geological genetic map through the target stratification function, combinations of risk factors without causal relationships are restricted, reducing the impact of spurious correlation combinations on risk assessment results. Finally, by improving the stratified evidence embedding layer in the evidence neural network model, the stratification results and explanatory strengths of risk factors are transformed into stratified evidence responses, enabling the coordinated output of risk level, uncertainty, and risk source. Overall, the results show that this method can improve the accuracy, interpretability, and reliability of priority ranking in geological exploration risk assessment.

[0073] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A geological exploration risk assessment method based on big data analysis, characterized in that, Includes the following steps: S1. Aggregate multi-source big data on geological exploration in the target exploration area, and organize it into a dataset of exploration objects and a mapping relationship between the locations of historical geological risk events according to the exploration objects; S2. Configure risk factor fields based on the exploration object dataset, construct a geological risk factor matrix, and generate a geological genetic map based on the geological genetic relationship between the risk factor fields. S3. Input the risk factors in the geological risk factor matrix into the improved optimal parameter geographic detector algorithm, construct the event projection correction function, embed the location mapping relationship of historical geological risk events into the candidate stratification process, construct the target stratification function, introduce the geological gene map to impose causal constraints on the combination of risk factors, and obtain the risk factor stratification results and the risk factor interpretation intensity. S4. Analyze the risk control relationship of the risk factor stratification results, identify the risk factors whose explanatory strength meets the control screening conditions as the main control risk factors, and identify the combination of risk factors along the genetic inheritance path in the geological genetic map to form the risk interpretation results. S5. Rearrange the risk factor codes based on the risk factor stratification results and risk interpretation results to form a risk assessment feature sequence; S6. Input the risk assessment feature sequence into the improved evidence neural network model containing a hierarchical evidence embedding layer to obtain the risk evidence vector of the exploration object. S7. Determine the risk level and risk uncertainty based on the risk evidence vector of the exploration object, determine the source of risk in combination with the risk interpretation results, and backfill to form the geological exploration risk zoning results and exploration priority results. 2.The geological exploration risk assessment method based on big data analysis of claim 1, wherein, S1 specifically includes: S11. Project the boundary of the target exploration area, the layout of exploration operations, and the existing exploration control positions onto a unified spatial reference, delineate the set of exploration objects, and assign object numbers, object boundaries, and object type markers to each exploration object; S12. Perform source identification, spatial attribution identification, and time period attribution identification on multi-source big data of geological exploration, and link the data that falls within the object boundary and the data that has an operational control relationship with the exploration object to the corresponding exploration object respectively; S13. Perform coverage segmentation and object attribution assignment on data spanning multiple exploration objects, and encapsulate the data that has completed the attribution assignment under the same exploration object within the object to obtain the exploration object dataset; S14. Determine the location and impact range of historical geological risk events, and link them to the corresponding exploration objects to form a location mapping relationship of historical geological risk events. 3.The geological exploration risk assessment method based on big data analysis of claim 1, wherein, S2 specifically includes: S21. Classify the data items in the exploration object dataset according to the causes of geological structure, stratigraphy, abnormal response, hydrological environment, engineering disturbance and historical events, and configure the risk factor field set. S22. Perform field repositioning on the data items in each exploration object, convert data items under the same causal category into corresponding risk factor field values, and configure data validity markers for missing fields. S23. Using the exploration object number as the matrix row index and the risk factor field as the matrix column index, the risk factor field values ​​and data validity markers are assembled in the same position to obtain the geological risk factor matrix. S24. The risk factor fields are processed into nodes according to their causal categories, and the geological causal relationship between the risk factor fields is converted into directed relationship edges to obtain the initial geological causal relationship diagram. S25. Verify the endpoints, directions, and duplicates of the directed edges in the initial geological genetic relationship diagram, retaining the directed edges with complete endpoints and consistent directions to obtain the geological genetic map. 4.The geological exploration risk assessment method based on big data analysis of claim 1, wherein, S3 specifically includes: S31. In the improved optimal parameter geographic detector algorithm, each risk factor field in the geological risk factor matrix is ​​expanded into a field value sequence, and candidate discrete segmentation is performed on the field value sequence to obtain a candidate stratification result set. S32. Perform interval assignment calibration on each candidate stratification result in the candidate stratification result set, label the stratification interval in each candidate stratification result as a candidate risk layer, and assign each exploration object to the corresponding candidate risk layer according to the corresponding risk factor value, and calculate the initial stratification score value of the candidate stratification result according to the factor detection rules of the geographic detector. S33. Project the location mapping relationship of historical geological risk events onto the candidate risk layer, count the density value of historical geological risk events in each candidate risk layer, and calculate the event density difference value between adjacent candidate risk layers. S34. Construct an event projection correction function. The event projection correction function takes the initial stratification score value, the historical geological risk event density value and the event density difference value as input. It performs score retention on the candidate stratification results where the historical geological risk events are concentrated in the high-risk layer, and performs score reduction on the candidate stratification results where the distribution of historical geological risk events is inconsistent with the candidate risk level. It outputs the event-corrected stratification score value. S35. Convert the geological genetic map into a risk factor causal adjacency table, configure effective causal adjacency values ​​for risk factor combinations with geological genetic inheritance relationships, and configure invalid causal adjacency values ​​for risk factor combinations without geological genetic inheritance relationships. S36. Construct a target stratification function. The target stratification function takes the event-corrected stratification score and the risk factor causal adjacency value as input. It performs masking on the candidate stratification results corresponding to the invalid causal adjacency values, retains the candidate stratification results corresponding to the valid causal adjacency values, and selects the target stratification result from the retained candidate stratification results. S37. The stratification boundary, stratification number, and score value corresponding to the target stratification result are determined as the risk factor stratification result and the risk factor explanatory strength.

5. The geological exploration risk assessment method based on big data analysis according to claim 1, characterized in that, S4 specifically includes: S41. Assign intervals to each risk factor in the risk factor stratification results according to the stratification number, and assign the explanatory strength of the risk factor to the corresponding stratification interval to obtain the single-factor risk explanation record. S42. Sort the risk factors in the single-factor risk interpretation record according to their explanatory strength. Mark the risk factors whose explanatory strength meets the main control screening criteria and whose historical geological risk event density in the corresponding stratified interval meets the event concentration criteria as the main control risk factors. S43. Locate the causal path of the main control risk factors in the geological causal map, connect the main control risk factors with geological causal relationship to obtain the candidate risk factor combination. S44. Cross-over the risk factor stratification intervals in the candidate risk factor combination, classify the exploration objects falling into the same overlapping interval into the same combination risk unit, and calculate the combination explanation strength of the combination risk unit. S45. Compare the combined explanatory strength of the combined risk unit with the corresponding single-factor risk explanatory record, configure a combined risk-causing label for candidate risk-causing factors whose explanatory strength is enhanced after combination, and determine the combined risk-causing factors. S46. Encapsulate the main control risk factors, main control stratification intervals, combined risk factors, combined risk units, and geological origination paths into risk interpretation results.

6. The geological exploration risk assessment method based on big data analysis according to claim 1, characterized in that, S5 specifically includes: S51. Match the stratification number in the risk factor stratification result with the main control risk factor and combined risk-causing factor in the risk interpretation result, and configure the main control label, combined label and stratification label for the risk factor field in the geological risk factor matrix; S52. Arrange the risk factor fields with the configured master control tags first, and determine the field order from high to low according to the explanatory strength of the risk factors; S53. Map the risk factor fields of the configured combination markers to the directed relation edges in the geological genealogy, extract the starting risk factor fields and the ending risk factor fields corresponding to the same combination of risk factors, and arrange them adjacently in the order of starting risk factor fields first and ending risk factor fields last to form a combination field segment. S54. Risk factor fields without configured master control and combination labels are retained in the supplementary field section and listed after the master control and combination field sections; S55. Encapsulate the risk factor values, stratification numbers, risk factor interpretation strengths, main control markers, and combination markers corresponding to each exploration object in the same position according to the rearranged field order to form a risk assessment feature sequence.

7. The geological exploration risk assessment method based on big data analysis according to claim 1, characterized in that, The improved evidence neural network model specifically includes a risk feature embedding layer, a hierarchical evidence embedding layer, a risk latent representation mapping layer, and an evidence output layer: The risk feature embedding layer divides the risk assessment feature sequence into field feature units according to the order of risk factor fields. Each field feature unit includes risk factor code, hierarchical number, risk factor explanatory strength, master control label and combination label. The risk factor codes in the field feature units are expanded into one-dimensional factor code vectors, and the stratified number, main control label, and combined label are converted into stratified label vectors, main control label vectors, and combined label vectors, respectively. The one-dimensional factor encoding vector, hierarchical label vector, master control label vector and combined label vector are concatenated in a fixed order within the fields to obtain the field embedding vector, and the field embedding vectors are arranged in the order of the risk factor fields to form a field embedding sequence. The hierarchical evidence embedding layer performs field matching and hierarchical position matching between the field embedding vectors in the field embedding sequence and the risk factor hierarchical results. It then performs hierarchical anchoring and evidence response transformation on the matched field embedding vectors to generate a hierarchical evidence response sequence. The risk implicit representation mapping layer divides the hierarchical evidence response sequence into the main control evidence segment, the combined evidence segment, and the supplementary evidence segment according to the main control label and the combined label. The hierarchical evidence response vector in each evidence segment is accumulated dimension by dimension to obtain the main control evidence vector, the combined evidence vector, and the supplementary evidence vector. The main evidence vector, combined evidence vector, and supplementary evidence vector are concatenated in sequence and then processed by linear mapping to obtain the risk hidden representation of the exploration object; The evidence output layer maps the implicit risk representation of the exploration object to the evidence channels corresponding to low risk, medium risk, high risk and extremely high risk, and arranges the output values ​​of each evidence channel in order of risk level to obtain the risk evidence vector of the exploration object.

8. The geological exploration risk assessment method based on big data analysis according to claim 7, characterized in that, The step of matching the field embedding vectors in the field embedding sequence with the risk factor stratification results for field matching and stratification position matching, and performing stratification anchoring and evidence response transformation on the matched field embedding vectors to generate a stratified evidence response sequence specifically includes: The risk factor stratification results are split into field stratification records according to the risk factor field. Each field stratification record includes the risk factor field, stratification number, stratification boundary, and risk factor explanatory strength. A stratification anchor vector is configured for each stratification number, and a stratification anchor table is established. For the target field embedding vector in the field embedding sequence, extract the corresponding risk factor field and stratification number, match the risk factor field and stratification number with the stratification anchoring table, and locate the target stratification anchoring vector. The target field embedding vector and the target hierarchical anchoring vector are added dimension by dimension to generate a hierarchical alignment vector; The explanatory strength of the risk factors corresponding to the target field embedding vector is expanded into an explanatory strength correction vector according to the dimensions of the hierarchical alignment vector. The hierarchical alignment vector and the explanatory strength correction vector are then multiplied dimension by dimension to generate a hierarchical response vector. For the hierarchical response vectors that are valid for the master control label, the response is retained to obtain the master control hierarchical response vector; for the hierarchical response vectors that are valid for the combined label, the hierarchical response vectors corresponding to adjacent risk factor fields in the same combined risk factor are extracted and then spliced ​​dimension by dimension according to the causal direction to obtain the combined hierarchical response vector. Perform scaling reduction on hierarchical response vectors where both master and combined tags are invalid to obtain supplementary hierarchical response vectors; The master hierarchical response vector, combined hierarchical response vector, and supplementary hierarchical response vector are rearranged according to field order to generate a hierarchical evidence response sequence.

9. The geological exploration risk assessment method based on big data analysis according to claim 1, characterized in that, Specifically, S7 includes: S71. The channel assignment of each risk level evidence value in the risk evidence vector of the exploration object is determined, and the low-risk evidence value, medium-risk evidence value, high-risk evidence value and extremely high-risk evidence value are respectively assigned to the corresponding risk level channel. S72. Normalize the evidence values ​​in each risk level channel to obtain the risk attribution value corresponding to each risk level, and determine the risk level with the largest risk attribution value as the risk level of the exploration object. S73. Accumulate all risk level evidence values ​​in the risk evidence vector of the exploration object to obtain the total amount of risk evidence, and determine the risk uncertainty based on the total amount of risk evidence; S74. Match the main control risk factors and combined risk factors in the risk interpretation results with the risk factor stratification results corresponding to the exploration object, extract the main control risk factors that fall into the dominant stratification interval and the combined risk factors that generate combined enhancement, and configure risk source markers. S75. Backfill the risk level, risk uncertainty and risk source markers into the spatial unit of the target exploration area to obtain the geological exploration risk zoning results. Then, prioritize each exploration object in order of risk level from high to low, risk uncertainty from low to high, and the number of combined risk factors from many to few to obtain the exploration priority results.

Images