Semi-supervised geological sample batch intelligent labeling method and system of mineralization knowledge rule

Abstract

The application relates to the technical field of geological data processing, and particularly discloses a semi-supervised geological sample batch intelligent labeling method and system of ore-forming knowledge rules, which comprises the following steps: generating an ore-forming knowledge rule set containing ore-forming condition constraint relationships; identifying un-labeled samples based on a preset sample classification model, and constructing a pseudo-label sample set containing category labels and confidence; screening the pseudo-label sample set based on the ore-forming knowledge rule set, expanding the labeled sample set, re-training the sample classification model based on the expanded sample set, and generating regional labeling results based on the trained sample classification model. The pseudo-label samples generated by the sample classification model are screened through the ore-forming knowledge rule set, only the high-confidence pseudo-labels meeting the ore-forming condition constraint relationships are used to expand the labeled sample set, and therefore, the continuous propagation of false pseudo-labels in the model training process is avoided.

Description

Technical Field

[0001] This invention relates to the field of geological data processing technology, specifically a semi-supervised intelligent batch annotation method and system for geological samples based on mineralization knowledge rules. Background Technology

[0002] Mineral resource prediction and regional mineral exploration evaluation are important research areas in the field of geological exploration. Their core lies in utilizing multi-source geological information, such as borehole data, geochemical data, geophysical data, and remote sensing imagery, to identify mineralization types and analyze the metallogenic potential of target areas, thereby providing a basis for mineral exploration deployment. With the continuous enrichment of data acquisition methods and the accumulation of large amounts of geological data, regional intelligent annotation methods based on machine learning models are gradually becoming an important technical means to improve mineral exploration efficiency.

[0003] However, the core technical challenge of existing intelligent region labeling methods lies in how to improve the category recognition accuracy of unlabeled regions given a limited number of labeled samples. To address this issue, existing technologies typically employ semi-supervised learning. This involves training a sample classification model using labeled samples, then using the trained model to predict unlabeled samples. The prediction results are then added to the training set as pseudo-labels, thereby expanding the sample size and improving model performance.

[0004] However, the aforementioned methods, when expanding the sample set using pseudo-labels, primarily rely on model predictions for sample selection, lacking constraints on mineralization patterns. When the model makes incorrect predictions for unlabeled samples, the erroneous categories are introduced into the expanded sample set and accumulate during subsequent training, causing the sample classification model to gradually deviate from the actual mineralization patterns, resulting in a decrease in the reliability of regional labeling results. This raises a more detailed technical issue: how to effectively screen pseudo-labeled samples using the mineralization constraints between geological entities to reduce model bias caused by the propagation of erroneous pseudo-labels and improve the accuracy and stability of regional labeling results.

[0005] Therefore, how to use the constraints of mineralization conditions to screen pseudo-labeled samples in order to improve the reliability of the sample classification model training process and improve the accuracy of the region labeling results is the technical problem that the present invention aims to solve. Summary of the Invention

[0006] The purpose of this invention is to provide a semi-supervised intelligent batch annotation method and system for geological samples based on mineralization knowledge rules, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A semi-supervised intelligent batch annotation method for geological samples based on metallogenic knowledge rules, the method comprising: Collect multi-source geological data, extract geological entity attributes and entity relationships, establish connection paths between geological attribute nodes, and generate a set of metallogenic knowledge rules containing metallogenic condition constraints. Based on a pre-defined sample classification model, unlabeled samples are identified, predicted categories and their category confidence scores are generated, and a pseudo-label sample set containing category labels and confidence scores is constructed. The pseudo-label sample set is filtered based on the mineralization knowledge rule set, the labeled sample set is expanded based on the filtered pseudo-label sample set, and the sample classification model is retrained based on the expanded sample set. The trained sample classification model determines the category of the location to be predicted in the target region and generates region labeling results.

[0008] As a further aspect of the present invention: the steps of collecting multi-source geological data, extracting geological entity attributes and entity relationships, establishing connection paths between geological attribute nodes, and generating a set of metallogenic knowledge rules containing metallogenic condition constraints include: Collect borehole data, geochemical data, geophysical data, and remote sensing image data, and unify the spatial coordinates of data from different sources; Extract mineralization type, lithology type, structural type, alteration type and elemental anomaly information to generate corresponding geological entity attributes; Statistical analysis of the co-occurrence relationships of different geological entity attributes in the same spatial location to determine the correlation relationships between geological entities; Geological attribute nodes are constructed based on the geological entity attributes and entity relationships, and connection paths between nodes are established; Based on the mineralization results corresponding to different node combinations according to the connection path, the mineralization condition constraints between nodes are determined, and a mineralization knowledge rule set is generated.

[0009] As a further aspect of the present invention: the step of identifying unlabeled samples based on a preset sample classification model, generating predicted categories and their category confidence scores, and constructing a pseudo-labeled sample set containing category labels and confidence scores includes: Extract lithological, structural, alteration, and elemental anomaly features from unlabeled samples to generate sample feature vectors; The sample feature vector is input into the first sample classification model to generate the first predicted category and the first category confidence. The sample feature vector is input into the second sample classification model to generate a second predicted category and a second category confidence score. Compare the first predicted category with the second predicted category to determine the candidate category label corresponding to the sample; Calculate the comprehensive category confidence based on the first category confidence and the second category confidence, and generate a pseudo-label sample set containing category labels and comprehensive category confidence.

[0010] As a further aspect of the present invention: the steps of filtering the pseudo-label sample set based on the mineralization knowledge rule set, expanding the labeled sample set based on the filtered pseudo-label sample set, and retraining the sample classification model based on the expanded sample set include: Read each pseudo-label sample in the pseudo-label sample set sequentially; The rule satisfaction degree of each pseudo-label sample that satisfies the metallogenic condition constraint relationship is calculated based on the metallogenic knowledge rule set; By comparing the rule satisfaction with a preset rule satisfaction threshold, pseudo-label samples whose rule satisfaction reaches the preset rule satisfaction threshold are filtered and extracted. The extracted pseudo-labeled samples are fused with the labeled sample set to generate an expanded sample set; The parameters of the sample classification model are recursively updated based on the expanded sample set to obtain the retrained sample classification model.

[0011] As a further aspect of the present invention: the step of determining the category of the location to be predicted in the target region based on the trained sample classification model and generating the region labeling result includes: Obtain the geological attribute data corresponding to the location to be predicted in the target area; Extract the sample feature vector corresponding to the location to be predicted based on the geological attribute data; The sample feature vector is input into the trained sample classification model to generate the class probability corresponding to the position to be predicted. By comparing the class probabilities corresponding to different categories, the target category corresponding to the location to be predicted is determined. Generate region labeling results for the target area based on the target category.

[0012] As a further aspect of the present invention: the mineralization knowledge rule set adopts an AST tree structure with recursive updates, and the recursive update process is as follows: In the process of calculating the rule satisfaction degree of each pseudo-label sample in satisfying the metallogenic condition constraint relationship based on the metallogenic knowledge rule set, the matching results between the pseudo-label samples and the metallogenic knowledge rules are statistically analyzed to identify rule conflict samples. Based on the rule conflict samples, the number of times each rule node is triggered, the number of times it is violated, and the degree of support are statistically analyzed to generate a rule contribution matrix; Calculate the error value corresponding to each rule node based on the rule contribution matrix; The AST tree structure is updated by adjusting the rule node weights, rule path connection strengths, and rule threshold parameters based on the error values. The updated AST tree structure is used to regenerate the mineralization knowledge rule set, which is then used in the next round of pseudo-label screening.

[0013] The present invention also provides a semi-supervised intelligent batch annotation system for geological samples based on mineralization knowledge rules, the system comprising: The rule set generation module is used to collect multi-source geological data, extract geological entity attributes and entity relationships, establish connection paths between geological attribute nodes, and generate a metallogenic knowledge rule set containing metallogenic condition constraints. The simulation annotation module is used to identify unlabeled samples based on a preset sample classification model, generate predicted categories and their category confidence scores, and construct a pseudo-label sample set containing category labels and confidence scores. The model training module is used to filter the pseudo-label sample set based on the mineralization knowledge rule set, expand the labeled sample set based on the filtered pseudo-label sample set, and retrain the sample classification model based on the expanded sample set. The model application module is used to determine the category of the location to be predicted in the target region based on the trained sample classification model and generate region labeling results.

[0014] As a further aspect of the present invention: the rule set generation module includes: The spatial unified unit is used to collect borehole data, geochemical data, geophysical data and remote sensing image data, and to unify the spatial coordinates of data from different sources; The entity attribute generation unit is used to extract mineralization type, lithology type, structural type, alteration type and elemental anomaly information, and generate corresponding geological entity attributes; The relationship determination unit is used to statistically analyze the co-occurrence relationships of different geological entity attributes in the same spatial location and determine the association relationships between geological entities. The path establishment unit is used to construct geological attribute nodes based on geological entity attributes and entity relationships, and to establish connection paths between nodes; The result combination unit is used to statistically analyze the mineralization results corresponding to different node combinations based on the connection path, determine the mineralization condition constraint relationship between nodes, and generate a mineralization knowledge rule set.

[0015] As a further aspect of the present invention: the simulation annotation module includes: The vector generation unit is used to extract lithological features, structural features, alteration features, and elemental anomaly features from unlabeled samples and generate sample feature vectors. The first prediction unit is used to input the sample feature vector into the first sample classification model to generate a first predicted category and a first category confidence level. The second prediction unit is used to input the sample feature vector into the second sample classification model to generate a second predicted category and a second category confidence level. The comparison unit is used to compare the first predicted category with the second predicted category to determine the candidate category label corresponding to the sample; The pseudo-label statistical unit is used to calculate the comprehensive category confidence based on the first category confidence and the second category confidence, and generate a pseudo-label sample set containing category labels and comprehensive category confidence.

[0016] As a further aspect of the present invention: the model training module includes: The sample reading unit is used to sequentially read each pseudo-label sample in the pseudo-label sample set; The satisfaction evaluation unit is used to calculate the rule satisfaction degree of each pseudo-label sample in satisfying the metallogenic condition constraint relationship based on the metallogenic knowledge rule set; The sample screening unit is used to compare the rule satisfaction with the preset rule satisfaction threshold and screen and extract pseudo-label samples whose rule satisfaction reaches the preset rule satisfaction threshold. The fusion unit is used to fuse the extracted pseudo-labeled samples with the labeled sample set to generate an expanded sample set; The recursive update unit is used to recursively update the parameters of the sample classification model based on the expanded sample set, so as to obtain the retrained sample classification model.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention extracts geological entity attributes and entity relationships from multi-source geological data to establish a metallogenic knowledge rule set containing metallogenic condition constraints. This metallogenic knowledge rule set is then used to screen pseudo-label samples generated by the sample classification model, using only highly reliable pseudo-labels that satisfy the metallogenic condition constraints to expand the labeled sample set, thereby preventing the continuous propagation of erroneous pseudo-labels during model training. Compared with existing methods that rely solely on model prediction results for sample expansion, this invention introduces geological metallogenic laws into the pseudo-label screening process, ensuring that the expanded samples simultaneously satisfy data feature laws and metallogenic condition constraints. This improves the reliability of the expanded samples, reduces model bias, enhances the ability of the sample classification model to represent complex metallogenic features, and thus improves the accuracy and stability of the target area category determination results, providing more reliable intelligent labeling results for regional mineral exploration evaluation. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention.

[0019] Figure 1The flowchart shows a semi-supervised intelligent batch annotation method for geological samples based on metallogenic knowledge rules.

[0020] Figure 2 The first sub-flowchart of the semi-supervised intelligent batch annotation method for geological samples based on metallogenic knowledge rules.

[0021] Figure 3 The second sub-flowchart of the semi-supervised intelligent batch annotation method for geological samples based on metallogenic knowledge rules.

[0022] Figure 4 The third sub-flowchart of the semi-supervised intelligent batch annotation method for geological samples based on metallogenic knowledge rules.

[0023] Figure 5 The fourth sub-flowchart of the semi-supervised intelligent batch annotation method for geological samples based on metallogenic knowledge rules.

[0024] Figure 6 This is a structural diagram of a semi-supervised intelligent batch annotation system for geological samples based on rules of mineralization knowledge.

[0025] Figure 7 This is a diagram illustrating the overall architecture of a semi-supervised intelligent batch annotation method for geological samples based on metallogenic knowledge rules. Detailed Implementation

[0026] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0027] Figure 1 The flowchart of a semi-supervised intelligent batch annotation method for geological samples based on metallogenic knowledge rules is shown below, combined with... Figure 7 In this embodiment of the invention, a semi-supervised intelligent batch annotation method for geological samples based on mineralization knowledge rules is provided, the method comprising: Step S100: Collect multi-source geological data, extract geological entity attributes and entity relationships, establish connection paths between geological attribute nodes, and generate a set of metallogenic knowledge rules containing metallogenic condition constraints. To address the challenge of algorithmically translating qualitative knowledge in the geological field, a dual-category system of entity constraint rules and numerical constraint rules is constructed, both employing a standardized production rule format. The formal definition is achieved using the rule = (Rule_ID, Condition_Set, Label_Type, Confidence, Applicable_Deposit_Type), where Condition_Set combines atomic conditions using logical operators, and Confidence is determined through a dual quantification method of expert scoring and verification of known mineral deposits. A matching rule parsing algorithm based on Abstract Syntax Tree (AST) is designed. Through geological data encoding (entity enumeration mapping, numerical normalization, spatial relationship calculation), rule AST generation (structured representation of root node-operator node-atomic condition node), recursive descent matching, and matching quantification, the formal rules are accurately adapted to multi-source geological data, solving the core problem in existing technologies where geological knowledge cannot directly participate in algorithmic decision-making.

[0028] Step S200: Identify unlabeled samples based on a preset sample classification model, generate predicted categories and their category confidence scores, and construct a pseudo-label sample set containing category labels and confidence scores; Two sample classification models are typically set up. These two models employ a progressive screening logic of "dual model consistency verification - mineralization rule matching - entropy constraint purification". First, pseudo-labels are generated in parallel for unlabeled samples using two lightweight models, MobileNet and ShuffleNet. Consistency samples are selected based on both probability consistency (Consistency_prob≥0.95) and type consistency, eliminating samples with prediction discrepancies and noise. Then, the mineralization rule parsing interface is called to retain samples with a comprehensive matching score of Match_Score≥0.8 and to correct the labels in conjunction with the rules. Finally, a high-quality pseudo-label set is generated by screening through pseudo-label entropy calculation (H(P)=-∑(P_prob×log2P_prob)) and dynamic threshold (0.2-0.4). At the same time, through spatial deduplication and conflict sample processing, the false positive rate of pseudo-labels is controlled below 12%, breaking through the accuracy bottleneck of existing single data-driven or knowledge-constrained screening.

[0029] Step S300: Filter the pseudo-label sample set based on the mineralization knowledge rule set, expand the labeled sample set based on the filtered pseudo-label sample set, and retrain the sample classification model based on the expanded sample set. Step S200 involves labeling unlabeled samples using existing models. Combined with the constructed mineralization knowledge rule set, some inappropriate labeling can be eliminated. For the remaining, better-labeled samples, knowledge distillation techniques are used to initialize the MobileNet and ShuffleNet dual-model architecture. Loss_Distill= λ×MSE(M1_Logit,M2_Logit)+(1-λ)×CrossEntropy(M1_Pred,L_Label) (λ takes values ​​of 0.3-0.5); complete T0-T10 rounds of initialization training to ensure the basic predictive ability of the model; in the iterative update stage, a multi-objective comprehensive loss function Loss_Total=Loss_Supervised(L)+α×Loss_Semi(P_High)+β×Loss_Consistency(M1,M2) is adopted, where α (0.4-0.6) is dynamically adapted with the quality of pseudo-labels, β is fixed at 0.2, and the credibility of mineralization rules is dynamically adjusted in combination with the feedback of annotation accuracy, forming a two-dimensional iterative closed loop of "model parameters-rule weights" to achieve accurate training in scenarios with scarce supervised samples and avoid overfitting or underfitting.

[0030] Step S400: Based on the trained sample classification model, determine the category of the location to be predicted in the target region and generate region labeling results; Steps S200 and S300 are cyclical processes. As the number of samples increases, the performance of the two models is continuously optimized. When the error rate is small enough, geological data (the process of constructing samples) is obtained for the predicted location in the area to be analyzed (target area). The geological data is then input into the trained sample classification model to generate the regional labeling results.

[0031] It should be noted that, in the technical solution of this invention, to adapt to the characteristics of strong spatial correlation and large volume of geological data, a 10km×10km spatial grid segmentation strategy is adopted, with a 500m overlap zone designed at the grid boundary to prevent the loss of boundary sample features; a data parallel framework is constructed based on ≥4 NVIDIA A100 GPUs, and the Horovod distributed training framework is used to achieve real-time synchronization of model parameters. Each GPU is responsible for the entire process of 1-2 data blocks (feature extraction, pseudo-label generation, rule matching); a pre-fetch + memory caching mechanism is used to optimize data loading, and data blocks are evenly split according to sample size (each block sample size ±5%) to ensure GPU load balance, and a single batch can process 10 6 The system can process 100,000 square kilometers of data in just 4 hours, balancing the efficiency of large-scale batch processing with the integrity of geological spatial features.

[0032] Figure 2 The first sub-flowchart of the semi-supervised intelligent batch annotation method for metallogenic knowledge rules includes the following steps: collecting multi-source geological data, extracting geological entity attributes and entity relationships, establishing connection paths between geological attribute nodes, and generating a set of metallogenic knowledge rules containing metallogenic condition constraints. Step S101: Collect borehole data, geochemical data, geophysical data, and remote sensing image data, and unify the spatial coordinates of data from different sources; Step S102: Extract mineralization type, lithology type, structural type, alteration type and elemental anomaly information to generate corresponding geological entity attributes; Step S103: Statistically analyze the co-occurrence relationships of different geological entity attributes in the same spatial location to determine the association relationships between geological entities; Step S104: Construct geological attribute nodes based on geological entity attributes and entity relationships, and establish connection paths between nodes; Step S105: Based on the connection path, statistically analyze the mineralization results corresponding to different node combinations, determine the mineralization condition constraints between nodes, and generate a mineralization knowledge rule set.

[0033] In one example of the technical solution of this invention, the process of constructing a set of rules for mineralization knowledge is described as follows: The above content explains the construction process of the metallogenic knowledge rule set. Based on existing metallogenic theoretical literature, typical deposit exploration reports, and geological expert experience summaries, core knowledge directly related to mineralization annotation is extracted, covering the relationship between "mineralization control elements, mineralization type, and data characteristics". For example, the mineralization control elements of porphyry copper deposits include "sericitization / chloritization alteration, fault structure, and Cu / Mo element anomaly", and the corresponding remote sensing and geochemical data characteristics are "hydroxyl anomaly value and element content threshold".

[0034] Furthermore, based on the constraint object and data type, mineralization knowledge is divided into two core rules, covering different mineralization annotation scenarios: 1) Entity Constraint Rules: Designed for the association relationships of geological entities (alteration type, structural type, stratigraphic lithology), used to constrain the entity combination conditions of mineralization samples. The core logic of the rules is "coexistence of multiple geological entities / associations increase the probability of mineralization / specific mineral type label", and is defined using a combination of "entity type + spatial relationship + logical operator".

[0035] Specific examples and parameter definitions: Entity Constraint Rule 1 (Porphyry Copper Deposit Ore Control Rule): R1: Alteration type ∈ {sericite, chloritization} ∧ Fault distance < 500m ∧ Stratigraphic lithology ∈ {granodiorite, quartz monzonite} ∧ Topographic slope ∈ [5°, 35°] ∧ Elevation ∈ [500m, 2000m], which indicates an increased mineralization probability; where "fault distance" is defined as the straight-line distance from the sample point to the nearest fault zone, calculated using GIS spatial analysis tools; "topographic slope" is extracted based on 30m resolution DEM data, using a slope calculation algorithm ( The Zevenbergen-Thorne method yields the elevation data, which is derived from the standardized DEM data. "∈" indicates "belongs to" and "∧" indicates "logical AND". The increase in mineralization probability is positively correlated with the completeness of rule matching. When the match is perfect, the probability increases by 30%-50%. When the terrain conditions do not match, the probability increase is reduced to less than 10% (porphyry copper deposits are mostly formed in low and medium hilly areas. If the slope is too gentle, it will easily lead to the leaching of mineralized alteration materials. If it is too steep, it will not be conducive to the accumulation of mineralized bodies. The elevation range is suitable for the mineralization elevation characteristics of most porphyry copper deposits).

[0036] Entity Constraint Rule 2 (Gold Deposit Control Rule): R2: Alteration type ∈ {Silicification, Pyrite} ∧ Within the ring structure ∧ Stratigraphic lithology is metamorphic rock ∧ Topographic slope ∈ [5°, 30°] ∧ Altitude ∈ [300m, 1800m] ∧ Topographic relief < 200m increases the probability of mineralization; where "within the ring structure" is defined as the sample point falling inside the boundary of the ring structure, and the ring structure is extracted by the Canny edge detection algorithm of remote sensing imagery (combined with ASTER 30m resolution remote sensing data, boundary error ≤ 20m); "Topographic slope" is based on 30m resolution DEM data, using Zevenbergen-T The horne method is used to calculate the elevation, which is taken from the DEM standardization results. The topographic relief is the difference in elevation between the highest and lowest points within a 1km radius of the sample point, calculated using GIS neighborhood analysis tools. The increase in mineralization probability is positively correlated with the degree of matching with topographic conditions. When the matching is perfect, the probability increases by 25%-40%. When a single topographic condition fails to meet the standard, the probability increase is reduced to less than 8%. (Gold deposits are mostly formed in low and medium mountain areas with gentle slopes. If the slope is too steep, the mineralized body is easily damaged by weathering and erosion. If it is too gentle, it is easily covered by sediments. Low-relief topography is conducive to the stable preservation of mineralized bodies and is suitable for the mineralization topographic features of most vein gold deposits and sedimentary metamorphic gold deposits.)

[0037] 2) Numerical Constraint Rules: These rules are designed with threshold conditions for quantitative data such as remote sensing, geochemical exploration, and geophysical exploration, and are used to constrain the range of numerical characteristics of mineralized samples. The core logic of the rules is "multi-source data values ​​satisfying threshold conditions up to a specific mineral type label," and they are defined using a combination of "data type + numerical threshold + logical operators."

[0038] Specific examples and parameter definitions: Numerical constraint rule 1 (porphyry copper deposit identification rule): R2: Remote sensing hydroxyl anomaly value > 0.8 ∧ geochemical Cu content > 200ppm ∧ geochemical Mo content > 5ppm, corresponding to porphyry copper deposit label; Among them, the remote sensing related parameters are refined as follows based on the characteristics of ASTER hyperspectral remote sensing data and typical remote sensing features of porphyry copper deposits: Data source and preprocessing: ASTER remote sensing image data (spatial resolution 30m, corresponding to the topographic DEM resolution mentioned above, ensuring spatial matching) are used. The preprocessing process includes atmospheric correction (FLAASH algorithm to eliminate water vapor and aerosol interference), topographic correction (C correction model to correct the influence of topographic shadows in low and medium hilly areas, adapting to the slope / altitude constraint scenario mentioned above), and band registration (error ≤ 0.5 pixels), laying the data foundation for anomaly extraction; Hydroxyl anomaly value calculation: Based on the ASTER image feature band combination, the 4 / 6 band ratio method is used (4 band wavelength 1.65μm, 6 ... The original ratio was calculated using a wavelength of 2.21 μm (all wavelengths sensitive to hydroxyl minerals), and then normalized using the minimum value removal method to obtain the hydroxyl anomaly value in the range of [0,1]. This ratio combination can accurately capture the remote sensing response of sericitization and chloritization alteration associated with porphyry copper deposits (corresponding to the alteration type of entity constraint rule 1). The threshold determination basis is as follows: in addition to statistically analyzing the distribution of typical porphyry copper deposit data, the spatial correlation verification between remote sensing anomalies and mineralization alteration zones is also combined. A threshold of 0.8 corresponds to a spatial overlap of ≥85% between the hydroxyl anomaly zone and the known porphyry copper alteration zone, which can effectively distinguish between mineralization alteration and vegetation and soil background interference (such as the hydroxyl anomaly value of weathering zone is generally ≤0.6). "ppm" is the numerical unit (10-6). The geochemical element content threshold is obtained by statistically analyzing geochemical data from 80 typical porphyry copper deposits, covering the numerical characteristics of more than 85% of known deposits, and forming a collaborative judgment logic with the remote sensing anomaly threshold.

[0039] Numerical constraint rule 2 (lithium ore identification rule): R3: Remote sensing Al-OH absorption depth > 0.15 nm ∧ Geochemical Li element content > 150 ppm ∧ Geophysical resistivity > 500 Ω·m, corresponding to spodumene ore label; Among them, the remote sensing related parameters are refined as follows based on the characteristics of hyperspectral remote sensing data and typical remote sensing responses of spodumene ore: Data source and preprocessing: Hyperion hyperspectral remote sensing data (spatial resolution 30 m, consistent with the resolution of DEM and ASTER data mentioned above, ensuring spatial matching; spectral resolution 10 nm, covering the 400 nm-2500 nm band), and the preprocessing process is synchronized with the previous steps. Standard: First, the FLAASH atmospheric correction algorithm is used to eliminate water vapor and aerosol interference. Then, the C-correction model is used to correct the topographic shadows in low and medium mountain areas (adapting to the topographic features of spodumene deposits, which are often distributed in low and medium mountain and hilly areas). Finally, band registration (error ≤ 0.5 pixels) and bad band removal (removing strong absorption bands above 2500nm) are performed to lay the data foundation for absorption depth calculation. Al-OH absorption depth calculation: Spodumene contains characteristic Al-OH groups, and its absorption peaks are concentrated in the wavelength range of 2.2μm-2.3μm. The "continuum removal method" is used to accurately extract the absorption depth: first, a linear continuous curve is fitted to this band range. The baseline is used to calculate the difference between the spectral reflectance of each pixel at the absorption peak position (2.25 μm) and the baseline of the continuum. This difference is the Al-OH absorption depth, ranging from [0, 0.5 nm]. This method can effectively shield the spectral interference of gangue minerals such as feldspar and quartz, and accurately capture the remote sensing response of spodumene deposits. The threshold is determined based on the statistical analysis of hyperspectral data from 50 typical spodumene deposits, combined with the correlation analysis between absorption depth and spodumene mineralization intensity. When the absorption depth is >0.15 nm, the spatial overlap with the spodumene mineralization body is ≥82%, while the Al-OH absorption depth of the surrounding rock (such as granite) is generally ≤0.0 nm. 8nm, which can effectively distinguish between mineralized and non-mineralized areas; multi-source data collaboration: remote sensing Al-OH absorption depth serves as a direct remote sensing indicator of spodumene mineralization, forming a collaborative judgment logic with geochemical Li element content (evidence of mineral composition) and geophysical resistivity (evidence of physical properties of mineralized bodies). At the same time, through spatial coordinate correlation, the spatial consistency (overlap ≥70%) between remote sensing anomaly areas and geochemical and geophysical anomaly areas is ensured, further improving the reliability of the rules; geophysical resistivity data comes from magnetotelluric sounding (MT) results, with a spatial resolution ≤50m. Through interpolation processing, it is adapted to the 30m remote sensing data resolution to ensure the spatial coupling of multi-source data.

[0040] Regarding the data structure of the above data, standardized production rules are used to represent all mineralization rules, with the following unified format: Rule=(Rule_ID,Condition_Set,Label_Type,Confidence,Applicable_Deposit_Type); The parameters are defined as follows: Rule_ID: A unique identifier for a rule, using the format "R + number" (e.g., R1, R2), used for the management and retrieval of the rule base; Condition_Set: A set of rule conditions, consisting of one or more atomic conditions combined by logical operators (∧: AND, ∨: OR, ¬: NOT). The format of the atomic condition is "data / entity type + constraint relationship + threshold / value range" (e.g., "alteration type ∈ {sericitization}" "Cu content > 200ppm"). Label_Type: The type of label output after a successful rule match, which is divided into two categories: "Increased Mineralization Probability" and "Specific Mineral Type Label". The specific mineral type label strictly follows the "Mineral Resources Classification Code" (GB / T17766-1999). Confidence: Rule confidence, with a value range of [0.7, 1.0]. It is quantified by "expert scoring + known deposit verification": 3-5 senior geological prospecting experts are invited to score the rationality of the rule (weight 0.4), and the final confidence is calculated by combining the matching accuracy of the rule in the known deposit dataset (weight 0.6). For example, if R1 has an expert score of 0.9 and a known deposit matching accuracy of 0.85, the final confidence = 0.9 × 0.4 + 0.85 × 0.6 = 0.87. Applicable_Deposit_Type: The type of deposit to which the rule applies, such as "porphyry copper deposit", "sedimentary bauxite", or "vein gold deposit", to avoid misuse of the rule across different scenarios.

[0041] Furthermore, all formalized rules are entered into the rule base, and a triple verification mechanism of "syntax verification + semantic verification + data verification" is adopted to ensure the validity of the rules: Syntax verification: Check whether the rule format conforms to the definition of production rule and whether the logical operators are used correctly; Semantic verification: Reviewed by geological experts to ensure that the rules conform to the metallogenic theory and have no logical contradictions (such as avoiding rules that violate metallogenic logic, such as "alteration type is silicification ∧ no tectonic development increases the probability of mineralization"); Data verification: Apply the rules to the labeled dataset of known deposits, remove rules with a matching accuracy of less than 60%, and optimize the parameters of rules with a matching accuracy of 60%-80% (such as adjusting thresholds and supplementing conditions).

[0042] Figure 3The second sub-flowchart of the semi-supervised intelligent batch annotation method for mineralization knowledge rules, which involves identifying unlabeled samples based on a preset sample classification model, generating predicted categories and their confidence levels, and constructing a pseudo-label sample set containing category labels and confidence levels, includes the following steps: Step S201: Extract lithological features, structural features, alteration features, and elemental anomaly features from unlabeled samples to generate sample feature vectors; Step S202: Input the sample feature vector into the first sample classification model to generate the first predicted category and the first category confidence. Step S203: Input the sample feature vector into the second sample classification model to generate the second predicted category and the second category confidence. Step S204: Compare the first predicted category with the second predicted category to determine the candidate category label corresponding to the sample; Step S205: Calculate the comprehensive category confidence based on the first category confidence and the second category confidence, and generate a pseudo-label sample set containing category labels and comprehensive category confidence.

[0043] The above content describes the application of two existing sample classification models to label unlabeled data. Both models are trained on existing labeled data (a small number). A small number of labeled geological samples (L, accounting for 5%-10% of the total sample size) serve as the supervised basis. The basic coding logic of the formal metallogenic rule base mentioned earlier is simultaneously loaded, and a dual lightweight model using MobileNet and ShuffleNet is initialized to build a basic predictive capability adapted to remote sensing and geochemical multi-source geological data. This design provides model support that aligns with metallogenic patterns for subsequent knowledge-enhanced pseudo-label generation. Furthermore, by complementing each other (capturing geological numerical features and spatial correlation features respectively), it avoids the misjudgment bias of a single model regarding geological noise (such as weathering zones and non-mineralized alteration), thus laying a solid foundation for accurate matching of subsequent metallogenic rules. The specific process is as follows: 1. Model selection and design basis: MobileNet and ShuffleNet were selected as the dual-student models (M1 and M2). The core selection logic revolved around the core requirements of "metallogenic knowledge-driven semi-supervised annotation, multi-source geological data adaptation, and large-scale batch processing," rather than selecting a general lightweight model. The specific basis is as follows: Lightweight architecture adapted for batch geological data processing: Both models have ≤5M parameters, resulting in 3-4 times faster inference speed compared to traditional CNNs (≥20M parameters), and are adaptable to 10... 6Parallel processing of multi-source geological samples (including 30m resolution ASTER remote sensing images, geochemical data, and topographic DEM data) provides architectural support for subsequent 10km×10km spatial block GPU parallel annotation, balancing processing efficiency and hardware compatibility. Feature extraction capabilities align with mineralization rule matching requirements: MobileNet's depthwise separable convolution excels at capturing local numerical features such as remote sensing hydroxyl anomalies and geochemical element content, accurately matching the numerical constraint rules mentioned above (e.g., R2 porphyry copper deposit remote sensing hydroxyl anomaly threshold, R3 lithium deposit Al-OH absorption depth); ShuffleNet's channel washing mechanism excels at mining spatial correlation features between alteration types, fault structures, and topographic slopes, adapting to spatial relationship determination in entity constraint rules (e.g., fault distance <500m, distribution within ring structures). The two complement each other, covering the two dimensions of geological data: "numerical features - spatial correlation features," laying a solid feature foundation for accurate matching of mineralization rules; Dual-branch output adapts to knowledge fusion annotation process: The model output layer is specially designed with "mineralization probability + mineral type classification" dual branches, which are fully aligned with the metallogenic rule base label system mentioned above. The mineralization probability branch outputs continuous values ​​in the [0,1] interval, providing data support for pseudo-label consistency verification and entropy constraint purification. The mineral type classification branch outputs discrete labels that conform to the "Mineral Resources Classification Code" (0=non-mineralized, 1=porphyry copper ore, 2=spodumene ore, etc.), which can be directly linked with the mineral type labels corresponding to entity / numerical constraint rules for correction, realizing the closed-loop optimization of labels of "model prediction - rule constraint", and strengthening the deep integration of metallogenic knowledge and semi-supervised annotation.

[0044] 2. Ore-forming knowledge-adaptive knowledge distillation initialization process: The core of this process is to rely on knowledge distillation technology to ensure that the dual models align with geological and mineralization patterns during the initialization phase. This involves simultaneously absorbing supervisory information from a small number of labeled samples and the geological semantic constraints of mineralization rules, thus building a basic predictive capability adapted to multi-source geological data. This lays a solid foundation for subsequent rule-driven pseudo-label selection. The specific process is as follows: Pre-training weight initialization and geological feature adaptation: The feature extraction layers of M1 (MobileNet) and M2 (ShuffleNet) are both initialized with ImageNet pre-trained weights. At the same time, the entity / numerical encoding mapping table of the metallogenic rule base mentioned above is loaded to perform preliminary fine-tuning of the weights. This ensures that the model focuses on the extraction of core geological features such as remote sensing alteration features (e.g., hydroxyl anomaly, Al-OH absorption depth), geochemical element content, and topographic slope from the initial stage, rather than general image features. The classification layer is randomly initialized, and the output dimension is strictly aligned with the metallogenic rule label system mentioned above (mineralization probability regression dimension + mineral type classification discrete dimension) to ensure compatibility with subsequent rule matching results.

[0045] Mineralization regularity-guided distillation training iterations (T0-T10): M1 is used as the core training model (emphasizing numerical feature fitting), and M2 is used as the auxiliary model (emphasizing spatial feature constraints). The learning of mineralization regularity is enhanced through the synergy of the two models. The loss function adopts a dual-objective design of "mineralization feature distillation loss + geological sample supervision loss", which both constrains the consistency of the two models' predictions and ensures that the model fits the real geological annotations. The formula is as follows: Loss_Distill= λ×MSE(M1_Logit,M2_Logit)+(1-λ)×CrossEntropy(M1_Pred,L_Label); The values ​​and basis for each parameter are as follows: λ: Distillation weight, with a value of 0.3-0.5 (determined through cross-validation of typical deposit samples, with 0.4 preferred). Its core function is to balance the "consistency of geological feature predictions between the two models" and the "supervision signal of the labeled samples", avoiding over-reliance on the auxiliary model and misjudging non-mineralization noise (such as weathering zone and vegetation interference), while ensuring the model's adaptability to mineralization rules.

[0046] M1_Logit / M2_Logit: The logits values ​​of the output layers M1 and M2. The focus is on capturing the intermediate predicted distribution of the dual models for core mineralization features such as "remote sensing alteration-tectonic spatial correlation-element anomaly", ensuring that the distillation process focuses on geological semantics rather than general features.

[0047] L_Label: The true label for sample L, strictly conforming to the mineralization rule label system mentioned above. It includes "mineralization probability regression label" (corresponding to the mineralization probability increase in the rule) and "mineral type classification label" (corresponding to specific mineral types in the R1-R3 rules), respectively adapting to the dual tasks of regression and classification, and strengthening the geological relevance of the monitoring signal.

[0048] MSE: Mean Squared Error, a core constraint ensuring the consistency of the two models' predictions of mineralization features, especially ensuring that the prediction biases of key numerical features such as hydroxyl anomalies and fracture distances, as well as spatial correlation features, are controllable; CrossEntropy: Cross-entropy loss, supervises the accurate fitting of M1 to real geological annotations, and avoids the model deviating from the mineralization regularity.

[0049] Iteration termination condition (geological adaptability guided): When the number of iterations reaches 10 rounds (T10), or the monitoring loss (CrossEntropy) decreases by ≤0.001 for 3 consecutive rounds, and the model's initial matching accuracy to the core mineralization rules (such as R1 porphyry copper ore control rules and R3 lithium ore determination rules) is ≥70%, the initialization is stopped, and the model weights of M1 and M2 at this time are saved as initial weights to ensure that the initialized model has both basic prediction capabilities and has been initially adapted to the mineralization rule logic.

[0050] 3. Post-initialization verification (metallogenic knowledge adaptability guidance): This validation is not a general model accuracy check; its core is to simultaneously validate the "basic predictive ability + mineralization rule adaptability" of the initial model. This ensures that the model meets both accuracy requirements and conforms to geological mineralization laws, providing a reliable foundation for subsequent knowledge-enhanced pseudo-label screening. The specific process is as follows: ① Validation set hierarchical construction: Based on the labeled sample L, sampling is performed hierarchically according to "mineral type (porphyry copper deposit / spodumene deposit / gold deposit) and mineralization type (tectonic control of mineralization / stratigraphic control of mineralization)". 80% is divided into training set and 20% into validation set to ensure that the validation set covers different mineralization scenarios and avoids validation bias caused by a single mineral type / mineralization type. The validation set must contain full-dimensional features such as remote sensing alteration, topographic slope, and geochemical elements, and be completely consistent with the data format in the rule modeling stage. ② Dual-dimensional verification indicators: In addition to the basic accuracy indicators, a new mineralization rule matching consistency indicator is added, forming a triple verification standard—basic accuracy: the accuracy of the mineralization probability prediction in the verification set is ≥75%, and the F1-score of the mineral type classification is ≥0.7 (weighted calculation for core mineral types); rule adaptability: the consistency between the model prediction results and the core mineralization rules (R1-R3) mentioned above is ≥70%, that is, among the samples predicted by the model as mineralized, at least 70% can hit the entity / numerical constraint rules of the corresponding mineral type, ensuring that the model prediction conforms to the geological logic; ③ Parameter adjustment and re-initialization logic: If the verification criteria are not met, prioritize adjusting the distillation weight λ (±0.1, adjusted specifically according to the type of deviation) — if the rule adaptability is not up to standard, lower λ (e.g., from 0.4 to 0.3) to strengthen the weight of geological monitoring signals in the labeled samples and improve the model's fit to the mineralization rule features; if the basic accuracy is not up to standard, raise λ (e.g., from 0.4 to 0.5) to enhance the consistency of the two models' predictions of geological features and reduce non-mineralization noise interference; after adjustment, re-execute the mineralization knowledge-adaptive distillation process until both indicators meet the standards, and record the optimal λ value and model weight as the initial benchmark for subsequent iterative optimization.

[0051] Figure 4 The third sub-flowchart of the semi-supervised intelligent batch annotation method for geological samples based on metallogenic knowledge rules includes the following steps: filtering the pseudo-label sample set based on the metallogenic knowledge rule set, expanding the labeled sample set based on the filtered pseudo-label sample set, and retraining the sample classification model based on the expanded sample set. Step S301: Read each pseudo-label sample in the pseudo-label sample set in sequence; Step S302: Calculate the rule satisfaction degree of each pseudo-label sample in terms of satisfying the metallogenic condition constraint relationship based on the metallogenic knowledge rule set; Step S303: Compare the rule satisfaction with the preset rule satisfaction threshold, and filter out pseudo-label samples whose rule satisfaction reaches the preset rule satisfaction threshold; Step S304: Merge the extracted pseudo-labeled samples with the labeled sample set to generate an expanded sample set; Step S305: Recursively update the parameters of the sample classification model based on the expanded sample set to obtain the retrained sample classification model.

[0052] In one example of the technical solution of this invention, a sample classification model optimization process based on a mineralization knowledge rule set is introduced. The sample classification model applied in step S200 is initially trained on a small amount of data, resulting in low accuracy. However, since this invention has constructed a mineralization knowledge rule set, it can optimize the results of the sample classification model, removing some unsuitable results. This process is then repeated, continuously refining the sample set, leading to increasingly higher performance of the sample classification model. The specific process is as follows: Through a triple-step verification process of "preliminary screening based on the consistency of geological features in dual models + rigid and precise screening based on metallogenic rules + purification of geological noise under entropy constraints," high-quality pseudo-label samples (P_High) that fit the metallogenic regularity are selected from a batch of unlabeled geological samples from multiple sources (U, including remote sensing alteration, geochemical elements, and topographic feature data). This process precisely removes non-mineralized noise samples such as those from weathering zones and vegetation cover, as well as samples with ambiguous features. The precise screening of metallogenic rules is one of the core innovations of the patent. It directly calls the output of the metallogenic rule parsing algorithm based on the Abstract Syntax Tree (AST) mentioned earlier, and combines it with the comprehensive matching degree (Match_Score) of the samples and the hit rule log (associated with entity / numerical constraint rules such as R1-R3), achieving deep coupling between metallogenic knowledge and multi-source geological data. This ensures that the pseudo-labels conform to both the model's prediction patterns and the geological metallogenic logic. Step 1: Initial screening based on consistency between the two models (first layer of denoising): The core of this step is to leverage the complementary predictive capabilities of the dual models for core mineralization features to perform initial denoising on a batch of unlabeled geological samples from multiple sources, selecting samples with strong predictive consensus. This lays a solid data foundation for subsequent screening based on rigid constraints of mineralization rules, while avoiding misjudgments of geological noise by the single model, ensuring labeling efficiency and basic reliability.

[0053] Parallel prediction using dual models: Unlabeled samples U, containing remote sensing alteration features (hydroxyl anomaly, Al-OH absorption depth), geochemical element content, and topographic factors (slope, elevation), are optimized for batch processing based on geological data processing efficiency (10 samples per batch). 4(Each sample is used to adapt to GPU memory load and geological data IO efficiency) as input to initialize M1 and M2. The two models output pseudo-labels P1 (M1 prediction, focusing on numerical feature fitting) and P2 (M2 prediction, focusing on spatial feature capture) respectively, based on the characteristics of the geological data. Each pseudo-label contains "mineralization probability (P1_prob / P2_prob, corresponding to the mineralization probability increase dimension in the mineralization rule) and mineral type classification (P1_type / P2_type, strictly aligned with the "Mineral Resources Classification Code" and the R1-R3 rule mineral type system)", ensuring seamless connection between the output and subsequent rules.

[0054] Consistency Calculation and Screening: A two-dimensional consistency index, Consistency, is defined based on geological annotation requirements to accurately quantify the consensus between the two models' predictions. Probability consistency: Consistency_prob=1-|P1_prob-P2_prob|, with a value range of [0,1]. This constraint focuses on the prediction deviation of the two models for mineralization probability, especially the fitting consistency of core mineralization numerical features such as remote sensing anomalies and geochemical element content, to avoid deviations caused by the sensitivity of a single model to numerical fluctuations. Type consistency: Consistency_type=1 (P1_type==P2_type) / 0 (otherwise), focusing on the accuracy of mineral classification, ensuring that the mineral type prediction of the screened samples has no fundamental conflict with the mineralization rule labeling system, and reducing the cost of subsequent rule correction.

[0055] Screening criteria: Only samples that simultaneously satisfy "Consistency_prob≥1-ε (ε=0.05)" and "Consistency_type=1" are retained to obtain the consistent sample set U_Consist. The value of ε=0.05 was determined through cross-validation of 50 typical mineral deposits (porphyry copper deposits and spodumene deposits), ensuring that more than 95% of the real mineralized samples are retained. At the same time, non-mineralized noise samples with large prediction discrepancies caused by weathering zone interference and vegetation cover are efficiently removed, balancing sample retention rate and noise reduction effect.

[0056] Step 2: Precise Screening of Ore-forming Rules (Knowledge Constraint Layer): This step, as the core carrier of knowledge-driven annotation, deeply reuses the AST parsing algorithm and rule base resources of the mineralization knowledge rule module mentioned above. Through the logic of "precise rule matching - rigid threshold screening - collaborative label correction", it transforms qualitative mineralization laws into rigid constraints for pseudo-label screening, solves the pain point of existing technologies that "data-driven approaches are prone to deviating from geological logic", and achieves deep coupling between model prediction and mineralization laws.

[0057] Rule matching call: Input the standardized coding features (entity enumeration coding, numerical normalization results, spatial relationship calculation values, fully aligned with the coding format in the rule modeling stage) of each sample in U_Consist into the AST rule parsing interface, and complete the conditional matching between the sample and the mineralization rule base (entity / numerical constraint rules such as R1-R3) based on the recursive descent matching algorithm. Output the "comprehensive matching degree (Match_Score), hit rule list, and matching log" of each sample. The comprehensive matching degree is calculated by weighting the credibility of the hit rules, which directly relates to the strength of the support of the rules for the mineralization attributes of the sample.

[0058] Screening criteria: Based on calibration using known deposit validation data, samples with a comprehensive matching score (Match_Score) ≥ 0.8 are retained to generate a rule-matching sample set U_Rule. This threshold is consistent with the effective hit threshold of the rule parsing algorithm, ensuring that over 90% of the samples in U_Rule conform to the mineralization logic of typical minerals such as porphyry copper deposits and spodumene deposits, accurately eliminating "pseudo-consistent" noise samples that are consistent with model predictions but violate mineralization rules (such as samples without alteration anomalies but misclassified as mineralized by the model).

[0059] Collaborative Label Correction: A label fusion mechanism of "model prediction + rule constraint" is constructed to make targeted corrections to samples in U_Rule: If a specific mineral type rule is matched (such as R2 porphyry copper ore rule, R3 spodumene ore rule), the pseudo-label mineral type classification is forcibly corrected to the mineral type corresponding to the rule, eliminating the ambiguity of the mineral type predicted by the model; at the same time, the mineralization probability is updated to "0.6 × model prediction probability + 0.4 × comprehensive matching degree", of which the rule matching degree accounts for 40% of the weight, strengthening the constraint priority of mineralization knowledge on labels, ensuring that the labels not only conform to the data feature patterns, but also meet the geological mineralization logic, and updating the label correction log synchronously to support subsequent dynamic correction and traceability.

[0060] Step 3: Entropy Constraint Optimization (Precise Purification Layer): This step, as the final purification stage of pseudo-label screening, is based on a dual verification of model prediction confidence and mineralization rule constraints. It eliminates samples with high prediction uncertainty and abnormal spatial distribution, ensuring that the final pseudo-label set possesses both data reliability and conforms to the spatial distribution patterns of mineralized bodies. This provides high-quality data support for subsequent iterative optimization. The detailed process is as follows: Pseudo-label entropy calculation (geological semantics oriented): For the U_Rule samples screened by mineralization rules, the pseudo-label entropy H(P) is calculated based on the "model-predicted mineralization probability (integrated with rule correction)", and the formula is H(P)=-∑(P_prob×Log2P_prob). The magnitude of entropy is directly related to the two dimensions of "model prediction confidence + mineralization rule consistency" — the lower the entropy, the clearer the model's prediction of the sample's mineralization characteristics (remote sensing alteration, geochemical anomaly, etc.), the more stable the consistency with the mineralization rules, and the smaller the label uncertainty; conversely, there may be problems such as blurred geological features (e.g., samples at the junction of alteration and weathering zone), marginal rule matching and other issues.

[0061] Dynamic threshold screening (adapted to mineralization scenarios): An entropy threshold H_Threshold is set, whose initial value 0.3 is determined by cross-validation of 50 groups of typical deposit samples (corresponding to the prediction accuracy of mineralization samples ≥ 92%). In each subsequent iteration, the threshold is dynamically adjusted according to the two indicators of "mineral species classification F1-score + rule matching consistency" on the verification set: if both indicators are improved simultaneously, the threshold is lowered by 0.05 (to strengthen purification and screen more accurate mineralized samples); if any indicator decreases, the threshold is increased by 0.03 (to retain more potential mineralized samples and avoid accidental deletion). The threshold is strictly controlled in the range of 0.2-0.4 (this range can balance "purification accuracy" and "sample retention rate", and adapt to differences in mineralization density in different metallogenic belts). Finally, samples with H(P)<H_Threshold are screened to form the preliminary high-quality pseudo-label set P_High.

[0062] Sample deduplication and abnormal conflict handling (conforming to geological spatial laws): Double post-processing is performed on P_High to completely eliminate the interference of abnormal samples: First, spatial deduplication: the region is divided into 100m×100m spatial grids (adapting to the accuracy of 30m resolution remote sensing and DEM data, and conforming to the minimum extension range of small mineralized bodies), only the sample with the highest mineralization probability is retained in each grid, which avoids model training deviation caused by repeated labeling in the same mineralized area; Second, rule conflict elimination: abnormal samples with "difference between mineralization probability and comprehensive matching degree > 0.3" are mainly eliminated — most of such samples are "model prediction meets the standard but rule matching is weak" (e.g., no alteration anomaly but misjudged as mineralization by the model), which violates the mineralization logic. After elimination, the false positive rate of pseudo-labels can be further reduced by 3-5%, ensuring that P_High samples fully meet the dual requirements of "data characteristics + geological rules".

[0063] The focus of the above content is the sample screening and update process. After the sample update is completed, labeled samples and pseudo-labeled samples are obtained. High-quality pseudo-labeled samples (P_High) that have been verified by mineralization rules and purified by entropy constraints are deeply integrated with the original labeled samples. The sample weights are dynamically adjusted based on the credibility of mineralization rules, and the fitting ability of the dual models to geological mineralization features is iteratively optimized. At the same time, combined with the spatial correlation of geological data, a spatial block GPU parallel strategy adapted to the distribution law of mineralization bodies is adopted to realize large-scale batch labeling of multi-source geological samples (remote sensing / geochemical exploration / topography). This not only consolidates the labeling accuracy through knowledge linkage, but also improves the processing efficiency through parallel architecture, perfectly meeting the core requirements of AI mineral exploration for samples that are "high-precision, large-volume, and fast". Specifically, relying on the dynamic constraints of mineralization knowledge rules, the linkage and iteration of "high-quality pseudo-labels - dual models - rule base" are realized. This not only balances the supervision and semi-supervision signals through multi-objective loss, but also strengthens the guidance of geological mineralization laws on model training, ensuring that the iterative model is more accurately adapted to the characteristics of multi-source geological data, and the labeling results have both data reliability and geological rationality.

[0064] Construction of the extended training set (hierarchical fusion of geological samples): High-quality pseudo-label samples P_High, which have been validated by mineralization rules and purified by entropy constraints, are deeply fused with the original manually labeled samples L to construct an extended training set adapted to geological scenarios. During fusion, the principle of "geological semantic alignment" is strictly followed—P_High samples must be accompanied by rule matching logs (associating with core mineralization rules such as R1-R3), and their label weight is set to 0.8 (lower than the original labeled sample weight of 1.0). This leverages the value of batch expansion of pseudo-labels while reducing the impact of potential geological biases of pseudo-labels (such as errors in rule edge matching samples) on the model through weight differences. When splitting the new training set, validation set, and test set, hierarchical sampling based on "mineral type + mineralization type" is adopted (covering typical mineral types such as porphyry copper deposits and spodumene deposits, as well as scenarios such as tectonic-controlled ore and stratigraphic-controlled ore). The ratio is strictly maintained at 7:2:1 to avoid overfitting of the model due to an excessively high proportion of samples from a single mineral type / mineralization scenario, ensuring that the dataset meets the labeling needs of AI mineral exploration across multiple scenarios.

[0065] Dual-model collaborative update (ore-forming feature-driven optimization): An "alternating training + multi-objective loss" model is used to update M1 (emphasizing numerical features) and M2 (emphasizing spatial features). The loss function Loss_Total deeply integrates geological annotation characteristics and ore-forming knowledge constraints, accurately balancing supervised signals, semi-supervised signals, and the consistency of the two models. The formula is as follows: Loss_Total=Loss_Supervised(L)+α×Loss_Semi(P_High)+β×Loss_Consistency(M1,M2) Each loss item is defined in detail based on the patent subject matter, and the parameters are designed accordingly: Loss_Supervised: The supervision loss of the original labeled samples is a weighted sum of "cross-entropy loss (mineral type classification task) + MSE (mineralization probability regression task)" with a weight ratio of 1:1. It is fully compatible with the dual-dimensional labeling system of "mineral type label + mineralization probability" in the mineralization rule base mentioned above, strengthens the guidance of real geological labels on the model, and ensures that the model fitting direction is close to the mineralization law.

[0066] Loss_Semi: Semi-supervised loss for pseudo-labeled samples, using label smoothing cross-entropy (smoothing coefficient 0.1), specifically designed for geological noise characteristics—by moderately smoothing the distribution of pseudo-labels, it reduces the impact of pseudo-label errors caused by non-mineralization interference such as weathering zones and vegetation cover on model training, while preserving the core feature differences under the constraints of mineralization rules.

[0067] Loss_Consistency: The consistency loss between the two models is calculated using MSE to determine the difference in mineralization probability predictions between the two models. This focuses on constraining the consistency of the models' predictions of core mineralization features such as remote sensing alteration anomalies, geochemical element content, and fault distance, thus avoiding geological misjudgments caused by feature extraction bias in a single model.

[0068] α: Semi-supervised loss weight, with a value of 0.4-0.6 (preferably 0.5), dynamically adjusted based on the quality of mineralization rule matching—if the mean comprehensive matching degree of mineralization rules for the P_High sample is ≥0.85 (i.e., the geological rationality of the pseudo-label is high), then α is 0.6 to enhance the expansion value of the pseudo-label; if the mean is <0.85, then α is 0.4 to reduce the interference of low-quality pseudo-labels, always taking geological rationality as the core of weight adjustment.

[0069] β: Consistency loss weight, fixed at 0.2, to balance the diversity of feature extraction and prediction consistency of the two models—it retains the advantages of M1 in extracting numerical features and M2 in extracting spatial features, while ensuring that the two models have consistent judgment logic for mineralization features, and avoids feature complementarity from turning into prediction conflict.

[0070] Iteration Termination Criteria (Dual Control of Geological Accuracy): Iteration termination must simultaneously meet the dual requirements of "data accuracy + geological rationality," controlling both model training convergence and preventing the model from deviating from mineralization patterns. Iteration stops when the number of iterations reaches 50 (T50), or when the F1-score for mineral classification on the test set improves by ≤0.005 for 5 consecutive iterations, the false positive rate of pseudo-labels is ≤15%, and the consistency between model predictions and mineralization rules is ≥80%. The final model weights and labeled sample set are then saved. The "rule matching consistency" metric is a patented design that ensures the model accurately fits the geological mineralization logic after iteration, rather than simply pursuing data fitting accuracy.

[0071] As a preferred embodiment of the technical solution of the present invention, a batch processing optimization process (large-scale annotation implementation adapted to mineralization scenarios) is provided: Based on the characteristics of geological data, such as strong spatial correlation, multi-source heterogeneity (remote sensing / geochemical exploration / topography), and the need for simultaneous execution of mineralization rule matching, a targeted optimization strategy was designed. Under the premise of ensuring that the constraints of mineralization knowledge are not weakened and the annotation accuracy is not reduced, the efficient parallel processing of large-scale geological samples is achieved, which is suitable for the "hourly" large-scale exploration area annotation needs of AI mineral exploration. At the same time, it is seamlessly connected with the pseudo-label screening and iterative optimization process mentioned above to ensure the synergy of technologies across the entire chain.

[0072] Spatial Block Strategy (Adaptation to Geological Spatial Characteristics): Considering the clustered distribution and limited extension range of mineralized bodies, and the spatial scale of 30m resolution remote sensing / DEM data, unlabeled samples U in batches are divided into 10km × 10km spatial grids, with each grid serving as an independent data block. This scale avoids the fragmentation of mineralized body features across grids (adapting to the 5-8km extension range of most small and medium-sized mineralized bodies), reduces the size of individual blocks, facilitates simultaneous loading of the mineralization rule base for matching calculations, and avoids rule matching deviations caused by interference from cross-grid sample features. A 500m overlap zone is designed at the grid boundaries to accurately cover scenarios where mineralized bodies may cross grid boundaries, preventing the loss of core mineralization features such as remote sensing alteration and fault distance of boundary samples, and ensuring the integrity of rule matching and labeling continuity of boundary samples.

[0073] GPU Parallel Implementation (Knowledge and Data Collaborative Parallelism): A mineralization knowledge-adaptive parallel processing framework is built based on ≥4 NVIDIA A100 GPUs. Employing a data parallel mode, each GPU independently handles the entire processing flow of 1-2 spatial data blocks. This encompasses three core stages: multi-source geological data feature extraction (M1 / M2 dual-model parallelism), mineralization rule AST parsing and matching, and pseudo-label generation and preliminary verification. This ensures that mineralization knowledge rule matching and model prediction are synchronized and parallel, without incurring additional time costs. Real-time synchronization of dual-model parameters is achieved through the Horovod distributed training framework, guaranteeing consistency in labeling standards for each data block. A single batch can stably process 10... 6 The multi-source geological sample processing efficiency is 3-4 times higher than that of single-card processing, completely solving the efficiency bottleneck of "disconnect between rule matching and parallel computing" in traditional batch processing.

[0074] Data loading optimization (geological data encoding adaptation): Addressing the IO characteristics of standardized geological data (including entity enumeration encoding, numerical normalization results, and spatial relationship data), a dual mechanism of "pre-reading + memory caching" is adopted. The geological data, after previous encoding processing, is cached in memory in advance to avoid lag caused by frequent disk data reads. This is particularly suitable for the rapid access of large-capacity data such as remote sensing images and geochemical element matrices. Simultaneously, data blocks are split according to the dual principles of "balanced sample size + balanced mineralization type," ensuring that the sample size fluctuation of each block is controlled within ±5%, and covering different mineralization scenarios such as tectonic mineralization and stratigraphic mineralization. This avoids load imbalance caused by a single GPU centrally processing samples of the same mineralization scenario, while ensuring uniform rule matching computation across GPUs, further improving parallel efficiency.

[0075] As a preferred embodiment of the technical solution of this invention, a dynamic correction process oriented towards the sample set is also introduced, providing a dynamic correction algorithm. This algorithm strictly adheres to the principles of "prioritizing geological laws, synergistic dual feedback, and controllable iterative convergence," employing a core process of "parallel dual feedback verification, priority-based execution, and iterative optimization convergence." Both feedback mechanisms deeply integrate mineralization knowledge rules and geological data characteristics, independently and accurately addressing one type of deviation while also forming a synergistic linkage: spatial feature feedback focuses on the geological attributes of natural mineralization aggregation, correcting spatial distribution anomalies; accuracy index feedback closely aligns with the credibility of mineralization rules, correcting rule-model matching deviations. When both types of anomalies are triggered simultaneously, spatial feature feedback correction is executed first (geological spatial laws have higher priority than statistical accuracy indicators), and then accuracy feedback correction is performed based on the spatially corrected samples, ensuring that the correction logic fully aligns with the spatial correlation and nonlinear mineralization laws of the mineralization system, adapting to multi-source geological data batch annotation scenarios.

[0076] The core logic of the above dynamic correction process is: As the core concluding link of the closed-loop architecture of "metallogenic knowledge - semi-supervised learning - annotation results" in this patent, the core objective is to rely on the metallogenic geological laws (spatial aggregation characteristics of mineralized bodies, metallogenic rule matching logic) to construct a dual mechanism of "metallogenic adaptive spatial feature feedback + metallogenic rule-guided accuracy index feedback". This mechanism performs targeted dynamic correction on the sample set generated by the semi-supervised batch annotation mentioned above, accurately eliminating annotation biases (such as spatial discrepancies that violate mineralization aggregation laws and false positive annotations caused by misjudgment of metallogenic rule matching). At the same time, it optimizes the credibility weight of the metallogenic rule base and the parameters of the two models (MobileNet / ShuffleNet) in reverse, forming a complete technical closed loop of "knowledge-driven annotation - geological feedback correction - rule model optimization - iterative re-annotation". This ensures that the large-scale batch annotation results not only conform to the data fitting laws, but also strictly fit the metallogenic geological logic, taking into account both accuracy stability and geological rationality.

[0077] The algorithm takes as input a batch of labeled sample sets (including labels, mineralization rule matching logs, and spatial coordinate information, consistent with the coding format in the rule modeling stage), a known deposit verification dataset (drill hole / reserve report calibration), and a mineralization rule base. The output is a corrected high-precision labeled sample set, an optimized mineralization rule base, and model fine-tuning weights, achieving a synergistic upgrade of labeling quality, knowledge rules, and model performance.

[0078] 1. Spatial autocorrelation feedback correction: The core of this mechanism relies on the inherent geological properties of mineralized bodies, namely "clustered distribution and non-random discreteness," and combines them with the spatial constraint logic in the mineralization rule system of this patent, such as "fault structure distance <500m" and "distribution within ring structure" in the entity constraint rules. By calculating the spatial autocorrelation index (Moran's I) of the labeled samples, it accurately identifies discrete labeled areas that violate the mineralization aggregation law, triggers the rematch of mineralization rules, corrects the "isolated labeled points" problem caused by the deviation of model prediction or initial rule matching, and ensures that the labeled results not only conform to the data law, but also fit the mineralization spatial logic.

[0079] 2. Moran's I Calculation Logic and Parameter Definition (Specific Adaptation Design for Mineralization Scenarios): ① Spatial weight matrix construction (adapting to the spatial correlation characteristics of mineralized bodies): The spatial weight matrix W is constructed using the "inverse distance weighting method," which is designed to fit the geological characteristic that "the closer the mineralized bodies are, the stronger the correlation," quantifying the mineralization correlation strength between samples. The formula is as follows: W_ij=1 / d_ij 2 (d_ij≠0); W_ij=0 (d_ij=0) Wherein, d_ij is the Euclidean distance (in meters) between the i-th sample and the j-th sample. Its value is strictly related to the spatial parameters in the mineralization rules mentioned above: only the correlation of samples within a radius of 5km is considered (based on the statistical determination of the extension range of 5-8km of typical small and medium-sized mineralization bodies such as porphyry copper deposits and spodumene deposits, which complements the 10km×10km spatial block scale mentioned above and avoids the break of mineralization correlation across blocks). If the range exceeds 5km, the weight is set to 0, which can effectively shield the interference of long-distance non-mineralized samples on the judgment of mineralization aggregation, and at the same time form a logical echo with the spatial constraints such as "fault distance" and "tectonic correlation" in the mineralization rules.

[0080] ② Moran's I Calculation (Adapted to Mineralization Probability Tag Characteristics): Moran's I has a value range of [-1, 1], specifically designed for the "mineralization probability + mineral classification" dual-tag system of this patent. It focuses on the spatial clustering calculation of the mineralization probability dimension. The calculation formula is as follows: I=[n×∑∑W_ij×(x_i-x - )×(x_j-x -)] / [S0×∑(x_i-x - ) 2 ].

[0081] Each parameter is defined in detail in conjunction with the annotation process: n is the total number of batch labeled samples. Low-confidence samples that have been screened out in the entropy constraint purification stage need to be removed. The calculation is only performed on high-quality labeled samples to ensure the reliability of clustering judgment. x_i is the mineralization probability value of the i-th sample (the final output value after correction by the mineralization rules, related to the Match_Score weighted result mentioned above), x - The average mineralization probability of all samples is calculated using a "mineral type weighted average" strategy (1.2 weight for core mineral types and 0.8 weight for non-core mineral types), which is suitable for multi-mineral type mixed annotation scenarios and avoids dilution of rare mineral type samples. S0 is the sum of all elements of the spatial weight matrix (∑∑W_ij), and its numerical distribution directly reflects the spatial density of the mineralized sample, providing geological semantic support for subsequent dispersion determination.

[0082] Dispersion Judgment Criteria (Metallogenic Regularity-Guided Threshold Design): Moran's I value is strongly correlated with mineralization clustering—the closer it is to 1, the stronger the spatial clustering of the sample, and the closer it is to the distribution pattern of mineralized body clusters; the closer it is to -1, the stronger the spatial dispersion, and the greater the probability of non-mineralized noise labeling or rule matching deviation. Combining known deposit verification data (statistical analysis of labeled samples covering 80 typical deposits), a threshold |I| < 0.2 is set to determine a high dispersion region: this threshold corresponds to an overlap of ≥ 88% between the high dispersion region and the non-mineralized noise region, which can accurately identify labeling deviations caused by weathering zone interference and misjudgment of isolated alteration points, and at the same time provide a preliminary judgment basis for subsequent metallogenic rule re-matching (threshold increased to 0.85), forming a progressive correction logic of "spatial feedback - knowledge constraint".

[0083] 3. Correction process for highly discrete regions (dual adaptation of mineralization rules linkage and geological spatial denoising): This process is the core implementation step of spatial autocorrelation feedback. It deeply reuses the patented mineralization rule base and AST parsing interface, and combines the inherent geological attributes of mineralization cluster distribution to achieve accurate correction and noise removal of discrete anomaly samples. At the same time, it connects to a semi-supervised annotation closed loop to ensure that the corrected samples not only conform to the mineralization regularity, but also adapt to the needs of subsequent model iterations. The specific steps are as follows: Anomaly area location (metallogenic spatial constraint guidance): Using GIS spatial visualization tools, locate the high-dispersion spatial range where |I|<0.2, and simultaneously overlay the structural and alteration spatial distribution data (such as fault zones and ring structure boundaries) from the entity constraint rules mentioned above. Extract all labeled samples (denoted as S_Abnormal) within this area and associate them with their original rule matching logs (including the hit entity / numerical constraint rules and Match_Score) to provide geological semantic support for subsequent rematching and avoid blind correction.

[0084] Metallogenic rule rematching (core rule reinforcement screening): The metallogenic rule parsing interface based on AST mentioned above is called to re-execute full rule matching on the S_Abnormal sample. The patented Match_Score weighted calculation logic (association rule credibility and condition matching completeness) is strictly reused, but the matching threshold is increased from 0.8 to 0.85. This threshold adjustment is determined based on the verification data of 80 typical deposits, which can improve the fit between the rematched sample and the mineralization aggregation pattern by more than 12%, strengthen the strictness of the geological rationality screening of discrete samples, and eliminate pseudo-rational samples that are consistent with the model prediction but violate the metallogenic rules.

[0085] Tag Correction and Removal (Dual-Tag System Adaptation Correction): Based on the re-matched Match_Score, differentiated processing is performed using the patented "mineralization probability + mineral type classification" dual-tag system, and the rule matching log is updated synchronously to support traceability. Match_Score≥0.85: The geological rationality of the sample is sufficient, the original mineralization probability and mineral type label are retained, and the update log is marked "the rule matching after spatial correction meets the standard", maintaining its weight (0.8) in the extended training set. 0.7≤Match_Score<0.85: The sample has a weak rule matching defect. The mineralization probability is reduced by 20% (corrected to "suspicious mineralization sample"). The mineral type label is retained but marked "rule matching needs to be verified". It is then assigned to the low-weight partition of the extended training set (weight 0.5) to reduce interference with model iteration. Match_Score<0.7: If the sample completely violates the mineralization rules (such as no alteration, isolated labels far from the structure), all labels are directly removed and the sample is assigned to the unlabeled sample set U. The sample then participates in the next round of the semi-supervised process of "dual model prediction - rule screening - entropy constraint purification" to avoid noise accumulation.

[0086] Spatial clustering denoising (adaptation to mineralization cluster characteristics): For the labeled samples, DBSCAN spatial clustering (eps=500m, min_samples=3) is performed to adapt to the mineralization density. The parameters are determined based on the cluster distribution density statistics of typical mineralizations such as porphyry copper deposits and spodumene deposits. 500m corresponds to the minimum clustering radius of most small and medium-sized mineralizations. min_samples=3 ensures that the clustering results conform to the geological logic of "multi-sample collaborative support for mineralization determination". Finally, isolated samples outside the cluster are removed. These samples have no surrounding mineralization samples or rules to support them. More than 90% of them are noise labels caused by weathering zones and isolated alteration points.

[0087] 4. Annotation accuracy feedback correction (dynamic optimization of mineralization rule weights + linkage of dual-model fine-tuning): This mechanism uses empirical data of known mineral deposits as a benchmark to construct a patent-specific verification system. By quantifying the annotation accuracy, it reverse-optimizes the credibility of the mineralization rule base and the parameters of the dual models (MobileNet / ShuffleNet) to solve the problem of "rule-model prediction deviation." At the same time, it strengthens the closed-loop linkage between knowledge rules and semi-supervised annotation to improve the reliability of subsequent batch annotation. The specific implementation is as follows: Validation set construction standards: Data source: Selected data of proven mineral deposits (including open-pit mines and borehole verification deposits) within the target exploration area, strictly covering core mineral types (porphyry copper deposits, spodumene deposits, vein gold deposits) and mineralization types (tectonic control mineralization, stratigraphic control mineralization), with a sample size of no less than 15% of the total number of batch annotation samples to ensure the representativeness of the verification results for the annotation accuracy of multiple scenarios.

[0088] Classification method: Stratified sampling is used based on "ore deposit type + metallogenic rule adaptability" to avoid bias caused by an excessive proportion of data of a single ore type or rule type. Ore deposit samples that can hit the core patent rules (R1-R3) are selected first, of which 80% are used as the accuracy verification set (V_Accuracy) to quantify the labeling accuracy; 20% are used as the rule weight optimization set (V_Weight) to specifically verify the actual adaptability of each metallogenic rule.

[0089] Labeling benchmark: The "mineralization boundary and mineral type" determined by the borehole core analysis results and the ore deposit reserve report are used as the real labels, strictly aligned with the patent labeling system. The accuracy of the mineralization boundary is controlled at 30m (consistent with the resolution of remote sensing and DEM data), which serves as a rigid benchmark for IoU calculation and mineral classification accuracy assessment.

[0090] Annotation accuracy calculation and threshold determination: Accuracy index definition: A dual-index evaluation is used, consisting of "classification accuracy" and "incoordination unit (IoU)" for mineralization boundary fit. The index calculation incorporates geological semantics. Specifically: Classification accuracy: Accuracy = (Number of true positive samples + Number of true negative samples) / Total number of samples in the validation set. True positive samples must simultaneously satisfy "model labeled mineralization, actual mineralization, and hit the corresponding mineralization rule". True negative samples must exclude "non-mineralized areas not covered by the rule" to avoid misjudging accuracy due to rule blind spots. Mineralization boundary fit: IoU = Model labeled mineralized area ∩ Actual mineralized area / Model labeled mineralized area ∪ Actual mineralized area. It focuses on evaluating the boundary accuracy of areas associated with mineralization rules (such as the area 500m around fault zones and within alteration zones), which is more in line with the actual needs of geological exploration. Anomaly threshold setting: Based on the accuracy quantification target (final Accuracy ≥ 0.88, IoU ≥ 0.75), the triggering condition is set: when Accuracy < 0.85 and IoU < 0.7, the annotation accuracy is judged to be substandard, and the rule weight adjustment and dual model fine-tuning are triggered simultaneously; when a single indicator is substandard, only the ore-forming rule re-matching is performed (reusing the AST parsing interface), without adjusting the weights and models, to avoid over-optimization leading to deviation from the ore-forming law.

[0091] Rule weight adjustment and model fine-tuning logic: Low-confidence rule identification: Based on the rule weight optimization set (V_Weight), the matching accuracy of each mineralization rule is calculated (the number of samples that match the real label after the rule is hit / the total number of hit samples of the rule). Rules with a matching accuracy of <70% are identified as low-confidence rules. Such rules often have problems such as unreasonable thresholds and missing conditions, and need to be optimized in a targeted manner to avoid misleading labeling.

[0092] Dynamic weight adjustment: Following the principle of "prioritizing geological rationality," the credibility of the rules is optimized to adapt to the management mechanism of this patent rule base. Low confidence rule: Adjust the confidence level according to the formula New_Confidence = Old_Confidence × 0.8~0.9. The lower the matching accuracy, the greater the adjustment (e.g., 60%-70% accuracy is adjusted to 0.9 times, <60% is adjusted to 0.8 times), and the minimum is not lower than 0.5 (to avoid the core ore control rules from becoming invalid, such as the R1 porphyry copper ore control rule); High-confidence rules: Rules with a matching accuracy of ≥85% (such as the R3 spodumene ore determination rule) have their confidence level increased by 0.05, with a maximum of 0.95, to strengthen their constraint priority in subsequent annotations. The updated rule base is synchronized to the AST parsing module.

[0093] Dual-model parameter fine-tuning: Based on the spatially corrected labeled sample set, the patent dual model (M1 / M2) is fine-tuned in a targeted manner, iterating for 5-8 rounds. The loss function is still Loss_Total, but the semi-supervised loss weight α is reduced to 0.3. By reducing the weight of pseudo-labels and strengthening the supervision role of real deposit labels (drilling data), the model's fitting deviation to non-mineralized noise is corrected. The fine-tuning termination condition is strictly aligned with the patent accuracy index: validation set accuracy ≥ 0.88 and IoU ≥ 0.75.

[0094] Rule base update feedback: The adjusted rule base is synchronized in real time to the rule parsing module mentioned above, serving as the core constraint basis for the next round of semi-supervised batch annotation, realizing a knowledge loop of "accuracy feedback - rule optimization - annotation upgrade", while recording rule adjustment logs to support geological experts in review and traceability.

[0095] Closed-loop iteration and exception conflict handling closed-loop linkage mechanism: (1) Closed-loop linkage: The corrected labeled sample set is replenished to the original labeled set L, and the updated rule base and dual model weights are used as inputs for the next round of semi-supervised batch labeling, deeply integrated into the whole process of "knowledge modeling-semi-supervised labeling-dynamic correction", forming a patent-specific self-iterative closed loop; (2) Handling of abnormal conflicts: When spatial autocorrelation anomaly and labeling accuracy anomaly are triggered at the same time, spatial autocorrelation correction is performed first—because the spatial aggregation of mineralized bodies is the core law of mineralization, its priority is higher than that of statistical accuracy indicators, so as to avoid violating geological logic by simply pursuing statistical accuracy. (3) Termination conditions: When the labeled sample meets all three conditions after two consecutive rounds of correction, the dynamic correction stops and the final high-precision sample set is output: Moran's I∈[0.2,0.8] (spatial clustering fits the mineralization law); Accuracy≥0.88 and IoU≥0.75 (precision meets the standard); the core mineralization rule (R1-R3) matching accuracy is ≥85% (knowledge constraints are effective).

[0096] Figure 5 The fourth sub-flowchart of the semi-supervised intelligent batch annotation method for mineralization knowledge rules, which involves determining the category of the location to be predicted in the target area based on the trained sample classification model and generating the area annotation results, includes: Step S401: Obtain the geological attribute data corresponding to the location to be predicted in the target area; Step S402: Extract the sample feature vector corresponding to the location to be predicted based on the geological attribute data; Step S403: Input the sample feature vector into the trained sample classification model to generate the class probability corresponding to the position to be predicted; Step S404: Compare the category probabilities corresponding to different categories to determine the target category corresponding to the location to be predicted; Step S405: Generate the region labeling results of the target region based on the target category.

[0097] In one example of the technical solution of the present invention, the final application process is described. Geological attribute data corresponding to the location to be predicted in the target area is obtained, sample feature vectors are generated in the same way as in step S200, the trained sample classification model is applied, and the corresponding prediction results are output. The prediction results include the predicted category and the corresponding category probability, which serve as the regional labeling results of the target area.

[0098] As a preferred embodiment of the technical solution of the present invention, the mineralization knowledge rule set adopts an AST tree structure with recursive updates, and the recursive update process is as follows: In the process of calculating the rule satisfaction degree of each pseudo-label sample in satisfying the metallogenic condition constraint relationship based on the metallogenic knowledge rule set, the matching results between the pseudo-label samples and the metallogenic knowledge rules are statistically analyzed to identify rule conflict samples. Based on the rule conflict samples, the number of times each rule node is triggered, the number of times it is violated, and the degree of support are statistically analyzed to generate a rule contribution matrix; Calculate the error value corresponding to each rule node based on the rule contribution matrix; The AST tree structure is updated by adjusting the rule node weights, rule path connection strengths, and rule threshold parameters based on the error values. The updated AST tree structure is used to regenerate the mineralization knowledge rule set, which is then used in the next round of pseudo-label screening.

[0099] As a preferred embodiment of the technical solution of this invention, the data structure of the metallogenic knowledge rule set is optimized, providing a recursively updated architecture. Specifically, a metallogenic rule parsing algorithm based on Abstract Syntax Tree (AST) is designed. Its core function is to transform formalized metallogenic rules into computer-executable matching logic, achieving accurate matching between geological data and metallogenic rules, and quantifying the degree of matching. The algorithm input is "a single geological sample data (including entity attributes and numerical features), and a metallogenic rule base," and the output is "a list of rules matched by the sample, the matching degree of each rule, and the comprehensive matching result." The specific implementation steps are as follows: 1. Geological data preprocessing and feature coding: The input geological sample data is standardized by encoding different types of geological data into a rule-matchable format, ensuring semantic alignment between the data and the rule conditions. The specific processing logic is as follows: 1) Entity data encoding: For entity attributes such as alteration type, stratigraphic lithology, and structural type, an "enumeration value mapping" encoding is adopted; for example, the alteration type "sericification" is mapped to enumeration value 1, "chloritization" is mapped to enumeration value 2, and "silicification" is mapped to enumeration value 3, generating an entity encoding table; the encoding table corresponds one-to-one with the entity values ​​in the rule base, ensuring that "alteration type ∈ {sericification, chloritization}" in the rule can be transformed into a matching condition of "entity code ∈ {1, 2}"; 2) Numerical data encoding: For numerical features such as remote sensing anomalies, geochemical element content, and structural distance, a "normalization + unit unification" encoding method is adopted. First, the numerical features are normalized to the [0,1] interval (normalization formula: Norm_Value=(Raw_Value-Min_Value) / (Max_Value-Min_Value), where Min_Value and Max_Value are the minimum and maximum values ​​of the feature in the target exploration area); second, the data units are unified (e.g., the "μg / g" of geochemical element content is uniformly converted to "ppm", 1μg / g=1ppm). 3) Spatial Relationship Data Encoding: For spatial relationship features such as "fracture distance" and "inside / outside ring structure", GIS spatial analysis algorithms are used to calculate and encode them; for example, "fracture distance" is obtained by calculating the shortest Euclidean distance between the sample point coordinates and the fracture zone polygon, and the unit is uniformly "meter (m)"; "inside / outside ring structure" is determined by the point-in-polygon algorithm and outputs a Boolean value (1 indicates inside, 0 indicates outside).

[0100] 2. Generation of AST for mineralization rules: Each rule in the mineralization rule base is traversed, and the rule's Condition_Set is transformed into an AST (Abstract Syntax Tree) to achieve a structured representation of the rule logic. The node definition and generation logic of the AST are as follows: 1) Node type definition: AST contains three core nodes—RootNode, Logical OperatorNode (such as ∧, ∨, ¬), and Atomic ConditionNode (such as "Alteration type ∈ {1,2}" "Cu content > 0.8 (after normalization)"); RootNode is the overall logical exit of the condition set, Logical OperatorNode is the intermediate node, and Atomic ConditionNode is the leaf node; 2) AST generation steps: Parse the Condition_Set of a single rule to obtain atomic conditions and logical operators; construct nodes in the order of "atomic conditions first, then logical operators", with the priority of logical operators following "¬>∧>∨"; with the root node as the top level, connect the logical operator nodes and atomic condition nodes in sequence to form a complete AST; for example, the Condition_Set of rule R1 "alteration type ∈ {1,2} ∧ fracture distance < 500m ∧ strata lithology ∈ {4,5}" (where 4 and 5 are enumerated values ​​of the corresponding lithology), the root node of the generated AST is "∧", which connects to three child nodes (two "∧" operator nodes and three atomic condition nodes), finally forming a tree structure of "root node ∧ to (∧ to (atomic condition 1 ∧ atomic condition 2)) ∧ atomic condition 3".

[0101] 3. AST matching based on recursive descent algorithm: A recursive descent algorithm is used to traverse the AST, performing atomic condition matching and logical operations node by node to achieve a complete match between the rules and geological sample data. The specific matching logic is as follows: 1) Atomic condition matching: Matching starts from the leaf nodes (atomic condition nodes) of the AST. The encoded geological sample data is compared with the atomic conditions, and the matching result of a single atomic condition is output (1 for matching, 0 for non-matching). For example, if the alteration type of the sample is encoded as 1 and the atomic condition is "alteration type ∈ {1,2}", the matching result is 1; if the fracture distance of the sample is 600m and the atomic condition is "fracture distance < 500m", the matching result is 0. 2) Logical operator matching: Recursively traverse the intermediate nodes (logical operator nodes) of the AST, perform the corresponding logical operation based on the matching results of the child nodes, and obtain the matching result of the operator node; for example, if the matching results of both child nodes of a "∧" node are 1, then the matching result of the node is 1; if the matching results of at least one of the two child nodes of a "∨" node are 1, then the matching result of the node is 1; if the matching result of the child node of a "¬" node is 1, then the matching result of the node is 0. 3) Overall rule matching: When the recursive traversal reaches the root node of the AST, the matching result of the root node is the overall matching result of a single rule with the sample (1 indicates that the rule is matched, and 0 indicates that it is not matched).

[0102] 4. Matching degree calculation and threshold determination: To avoid rule omissions or misjudgments caused by rigid "black and white" matching, a matching score (Match_Score) is introduced to quantify the degree of matching between the sample and the rule, and a threshold is used to filter valid matching results. The specific calculation and judgment logic is as follows: 1) Match score calculation formula: Match_Score=(∑(Atomic_Condition_Match×Atomic_Weight))×Rule_Confidence; where: Atomic_Condition_Match is the matching result of a single atomic condition (1 / 0); Atomic_Weight is the weight of a single atomic condition, set according to the degree of influence of the condition on mineralization. The weight of core ore-controlling conditions (such as "alteration type" and "elemental anomaly value") is set to 0.6-0.8, and the weight of auxiliary conditions (such as "stratigraphic lithology") is set to 0.2-0.4. The sum of the weights of all atomic conditions within a single rule is 1; Rule_Confidence is the credibility of the rule (quantified in the rule system construction stage). 2) Threshold Determination: By statistically analyzing the matching degree distribution between known mineralized and non-mineralized samples, the matching degree threshold was determined to be 0.8 using ROC curve analysis. When Match_Score ≥ 0.8, it is judged as "valid hit", and this rule can be used for subsequent pseudo-label screening. When 0.5 ≤ Match_Score < 0.8, it is judged as "suspicious hit", and needs to be combined with other rules for comprehensive judgment. When Match_Score < 0.5, it is judged as "miss", and the constraint of this rule on the current sample is excluded. 3) Multi-rule comprehensive matching: If a single sample hits multiple rules at the same time, the average matching degree of all valid hit rules is calculated as the "comprehensive matching degree" of the sample. The higher the comprehensive matching degree, the higher the degree to which the sample conforms to the mineralization rules, and the higher the priority of being judged as a high-quality pseudo-label.

[0103] 5. Output of parsing results: After parsing all rules for a single sample, a structured parsing result is output, including "sample ID, list of hit rules (including Rule_ID, Confidence), matching degree of each rule, overall matching degree, and matching result judgment (valid hit / suspicious hit / no hit)". This result will be directly input into the subsequent pseudo-label filtering module as the basis for geological semantic constraints.

[0104] Figure 6 The diagram shows the structural composition of a semi-supervised intelligent batch annotation system for geological samples based on metallogenic knowledge rules. In this embodiment of the invention, a semi-supervised intelligent batch annotation system 10 for geological samples based on metallogenic knowledge rules is provided. The system 10 includes: The rule set generation module 11 is used to collect multi-source geological data, extract geological entity attributes and entity associations, establish connection paths between geological attribute nodes, and generate a metallogenic knowledge rule set containing metallogenic condition constraints. The simulation annotation module 12 is used to identify unlabeled samples based on a preset sample classification model, generate predicted categories and their category confidence scores, and construct a pseudo-label sample set containing category labels and confidence scores. Model training module 13 is used to filter the pseudo-label sample set based on the mineralization knowledge rule set, expand the labeled sample set based on the filtered pseudo-label sample set, and retrain the sample classification model based on the expanded sample set. Model application module 14 is used to determine the category of the location to be predicted in the target region based on the trained sample classification model and generate region labeling results.

[0105] Furthermore, the rule set generation module 11 includes: The spatial unified unit is used to collect borehole data, geochemical data, geophysical data and remote sensing image data, and to unify the spatial coordinates of data from different sources; The entity attribute generation unit is used to extract mineralization type, lithology type, structural type, alteration type and elemental anomaly information, and generate corresponding geological entity attributes; The relationship determination unit is used to statistically analyze the co-occurrence relationships of different geological entity attributes in the same spatial location and determine the association relationships between geological entities. The path establishment unit is used to construct geological attribute nodes based on geological entity attributes and entity relationships, and to establish connection paths between nodes; The result combination unit is used to statistically analyze the mineralization results corresponding to different node combinations based on the connection path, determine the mineralization condition constraint relationship between nodes, and generate a mineralization knowledge rule set.

[0106] Specifically, the simulation annotation module 12 includes: The vector generation unit is used to extract lithological features, structural features, alteration features, and elemental anomaly features from unlabeled samples and generate sample feature vectors. The first prediction unit is used to input the sample feature vector into the first sample classification model to generate a first predicted category and a first category confidence level. The second prediction unit is used to input the sample feature vector into the second sample classification model to generate a second predicted category and a second category confidence level. The comparison unit is used to compare the first predicted category with the second predicted category to determine the candidate category label corresponding to the sample; The pseudo-label statistical unit is used to calculate the comprehensive category confidence based on the first category confidence and the second category confidence, and generate a pseudo-label sample set containing category labels and comprehensive category confidence.

[0107] Furthermore, the model training module 13 includes: The sample reading unit is used to sequentially read each pseudo-label sample in the pseudo-label sample set; The satisfaction evaluation unit is used to calculate the rule satisfaction degree of each pseudo-label sample in satisfying the metallogenic condition constraint relationship based on the metallogenic knowledge rule set; The sample screening unit is used to compare the rule satisfaction with the preset rule satisfaction threshold and screen and extract pseudo-label samples whose rule satisfaction reaches the preset rule satisfaction threshold. The fusion unit is used to fuse the extracted pseudo-labeled samples with the labeled sample set to generate an expanded sample set; The recursive update unit is used to recursively update the parameters of the sample classification model based on the expanded sample set, so as to obtain the retrained sample classification model.

[0108] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A semi-supervised intelligent batch annotation method for geological samples based on metallogenic knowledge rules, characterized in that, The method includes: Collect multi-source geological data, extract geological entity attributes and entity relationships, establish connection paths between geological attribute nodes, and generate a set of metallogenic knowledge rules containing metallogenic condition constraints. Based on a pre-defined sample classification model, unlabeled samples are identified, predicted categories and their category confidence scores are generated, and a pseudo-label sample set containing category labels and confidence scores is constructed. The pseudo-label sample set is filtered based on the mineralization knowledge rule set, the labeled sample set is expanded based on the filtered pseudo-label sample set, and the sample classification model is retrained based on the expanded sample set. The trained sample classification model determines the category of the location to be predicted in the target region and generates region labeling results.

2. The semi-supervised intelligent batch annotation method for geological samples based on metallogenic knowledge rules according to claim 1, characterized in that, The steps of collecting multi-source geological data, extracting geological entity attributes and entity relationships, establishing connection paths between geological attribute nodes, and generating a set of metallogenic knowledge rules containing metallogenic condition constraints include: Collect borehole data, geochemical data, geophysical data, and remote sensing image data, and unify the spatial coordinates of data from different sources; Extract mineralization type, lithology type, structural type, alteration type and elemental anomaly information to generate corresponding geological entity attributes; Statistical analysis of the co-occurrence relationships of different geological entity attributes in the same spatial location to determine the correlation relationships between geological entities; Geological attribute nodes are constructed based on the geological entity attributes and entity relationships, and connection paths between nodes are established; Based on the mineralization results corresponding to different node combinations according to the connection path, the mineralization condition constraints between nodes are determined, and a mineralization knowledge rule set is generated.

3. The semi-supervised intelligent batch annotation method for geological samples based on metallogenic knowledge rules according to claim 1, characterized in that, The steps of identifying unlabeled samples based on a preset sample classification model, generating predicted categories and their category confidence scores, and constructing a pseudo-labeled sample set containing category labels and confidence scores include: Extract lithological, structural, alteration, and elemental anomaly features from unlabeled samples to generate sample feature vectors; The sample feature vector is input into the first sample classification model to generate the first predicted category and the first category confidence. The sample feature vector is input into the second sample classification model to generate a second predicted category and a second category confidence score. Compare the first predicted category with the second predicted category to determine the candidate category label corresponding to the sample; Calculate the comprehensive category confidence based on the first category confidence and the second category confidence, and generate a pseudo-label sample set containing category labels and comprehensive category confidence.

4. The semi-supervised intelligent batch annotation method for geological samples based on metallogenic knowledge rules according to claim 1, characterized in that, The steps of filtering the pseudo-labeled sample set based on the mineralization knowledge rule set, expanding the labeled sample set based on the filtered pseudo-labeled sample set, and retraining the sample classification model based on the expanded sample set include: Read each pseudo-label sample in the pseudo-label sample set sequentially; The rule satisfaction degree of each pseudo-label sample that satisfies the metallogenic condition constraint relationship is calculated based on the metallogenic knowledge rule set; By comparing the rule satisfaction with a preset rule satisfaction threshold, pseudo-label samples whose rule satisfaction reaches the preset rule satisfaction threshold are filtered and extracted. The extracted pseudo-labeled samples are fused with the labeled sample set to generate an expanded sample set; The parameters of the sample classification model are recursively updated based on the expanded sample set to obtain the retrained sample classification model.

5. The semi-supervised intelligent batch annotation method for geological samples based on metallogenic knowledge rules according to claim 1, characterized in that, The steps for determining the category of the location to be predicted in the target region based on the trained sample classification model and generating region labeling results include: Obtain the geological attribute data corresponding to the location to be predicted in the target area; Extract the sample feature vector corresponding to the location to be predicted based on the geological attribute data; The sample feature vector is input into the trained sample classification model to generate the class probability corresponding to the position to be predicted. By comparing the class probabilities corresponding to different categories, the target category corresponding to the location to be predicted is determined. Generate region labeling results for the target area based on the target category.

6. The semi-supervised intelligent batch annotation method for geological samples based on metallogenic knowledge rules according to claim 1, characterized in that, The mineralization knowledge rule set adopts an AST tree structure with recursive updates. The recursive update process is as follows: In the process of calculating the rule satisfaction degree of each pseudo-label sample in satisfying the metallogenic condition constraint relationship based on the metallogenic knowledge rule set, the matching results between the pseudo-label samples and the metallogenic knowledge rules are statistically analyzed to identify rule conflict samples. Based on the rule conflict samples, the number of times each rule node is triggered, the number of times it is violated, and the degree of support are statistically analyzed to generate a rule contribution matrix; Calculate the error value corresponding to each rule node based on the rule contribution matrix; The AST tree structure is updated by adjusting the rule node weights, rule path connection strengths, and rule threshold parameters based on the error values. The updated AST tree structure is used to regenerate the mineralization knowledge rule set, which is then used in the next round of pseudo-label screening.

7. A semi-supervised intelligent batch annotation system for geological samples based on metallogenic knowledge rules, characterized in that, The system includes: The rule set generation module is used to collect multi-source geological data, extract geological entity attributes and entity relationships, establish connection paths between geological attribute nodes, and generate a metallogenic knowledge rule set containing metallogenic condition constraints. The simulation annotation module is used to identify unlabeled samples based on a preset sample classification model, generate predicted categories and their category confidence scores, and construct a pseudo-label sample set containing category labels and confidence scores. The model training module is used to filter the pseudo-label sample set based on the mineralization knowledge rule set, expand the labeled sample set based on the filtered pseudo-label sample set, and retrain the sample classification model based on the expanded sample set. The model application module is used to determine the category of the location to be predicted in the target region based on the trained sample classification model and generate region labeling results.

8. The semi-supervised intelligent batch annotation system for geological samples based on metallogenic knowledge rules according to claim 7, characterized in that, The rule set generation module includes: The spatial unified unit is used to collect borehole data, geochemical data, geophysical data and remote sensing image data, and to unify the spatial coordinates of data from different sources; The entity attribute generation unit is used to extract mineralization type, lithology type, structural type, alteration type and elemental anomaly information, and generate corresponding geological entity attributes; The relationship determination unit is used to statistically analyze the co-occurrence relationships of different geological entity attributes in the same spatial location and determine the association relationships between geological entities. The path establishment unit is used to construct geological attribute nodes based on geological entity attributes and entity relationships, and to establish connection paths between nodes; The result combination unit is used to statistically analyze the mineralization results corresponding to different node combinations based on the connection path, determine the mineralization condition constraint relationship between nodes, and generate a mineralization knowledge rule set.

9. The semi-supervised intelligent batch annotation system for geological samples based on metallogenic knowledge rules according to claim 7, characterized in that, The simulation annotation module includes: The vector generation unit is used to extract lithological features, structural features, alteration features, and elemental anomaly features from unlabeled samples and generate sample feature vectors. The first prediction unit is used to input the sample feature vector into the first sample classification model to generate a first predicted category and a first category confidence level. The second prediction unit is used to input the sample feature vector into the second sample classification model to generate a second predicted category and a second category confidence level. The comparison unit is used to compare the first predicted category with the second predicted category to determine the candidate category label corresponding to the sample; The pseudo-label statistical unit is used to calculate the comprehensive category confidence based on the first category confidence and the second category confidence, and generate a pseudo-label sample set containing category labels and comprehensive category confidence.

10. The semi-supervised intelligent batch annotation system for geological samples based on metallogenic knowledge rules according to claim 7, characterized in that, The model training module includes: The sample reading unit is used to sequentially read each pseudo-label sample in the pseudo-label sample set; The satisfaction evaluation unit is used to calculate the rule satisfaction degree of each pseudo-label sample in satisfying the metallogenic condition constraint relationship based on the metallogenic knowledge rule set; The sample screening unit is used to compare the rule satisfaction with the preset rule satisfaction threshold and screen and extract pseudo-label samples whose rule satisfaction reaches the preset rule satisfaction threshold. The fusion unit is used to fuse the extracted pseudo-labeled samples with the labeled sample set to generate an expanded sample set; The recursive update unit is used to recursively update the parameters of the sample classification model based on the expanded sample set, so as to obtain the retrained sample classification model.

Images