A Method and System for Intelligent Mineral Identification in Weak Information Areas Based on Transfer Learning

Abstract

This application provides a method and system for intelligent identification of minerals in weak information areas based on transfer learning, relating to the field of transfer learning technology. First, a source domain mineral sample set and a set of mineral samples to be identified in weak information areas are obtained. Then, geological correlation elements between the two are extracted and a cross-domain feature mapping is constructed. The original features of the minerals in the weak information areas are subjected to transfer adaptation processing to generate intermediate features of the minerals in the weak information areas that are adapted to the feature dimensions of the source domain. Next, sparse feature dimensions in the intermediate features are identified, and the corresponding feature distribution patterns of the source domain are used to complete the features, generating completed features. The completed features are input into a pre-trained transfer learning identification model for cross-domain identification, generating preliminary identification results. Finally, source domain samples with similar geological correlation elements are selected to determine the transfer deviation, correct the preliminary identification results, and generate the final identification results, effectively improving the accuracy and reliability of mineral identification in weak information areas.

Description

Technical Field

[0001] This application relates to the field of transfer learning technology, and more specifically, to a method and system for intelligent identification of mineral resources in weak information areas based on transfer learning. Background Technology

[0002] In the field of mineral resource exploration, accurately identifying mineral types in different areas is crucial for the rational planning of exploration work and improving resource development efficiency. Currently, traditional methods for mineral identification mainly rely on training models using a large amount of labeled sample data to classify and identify unknown samples. However, in actual exploration scenarios, there are many areas with weak information. Due to complex geological conditions and difficulties in data collection, the mineral samples obtained in these areas often lack clear mineral type labeling information, i.e., the mineral type is not labeled.

[0003] Existing mineral identification methods based on traditional machine learning face numerous challenges when processing mineral samples in weakly informative areas. Because these samples lack annotations, it's difficult to directly utilize them to build effective identification models. If models are trained using only labeled source domain samples (typically from areas with relatively clear geological conditions and well-collected data) and directly applied to weakly informative areas, the difference in geological features between the source and weakly informative regions leads to a significant drop in identification accuracy. Furthermore, the geological correlation elements in different regions are complex and diverse; simple feature mapping or model transfer cannot fully consider the correspondences and dynamic changes between these elements, failing to effectively address the problems of feature sparsity and transfer bias in mineral identification in weakly informative areas. This limits the application effectiveness and scope of mineral identification technology in these areas. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method and system for intelligent identification of mineral resources in weak information areas based on transfer learning.

[0005] According to a first aspect of this application, a method for intelligent identification of mineral resources in weak information areas based on transfer learning is provided, the method comprising:

[0006] Obtain a source domain mineral sample set and a weak information area mineral sample set to be identified. The source domain mineral sample set contains source domain mineral features with labeled mineral types, and the weak information area mineral sample set to be identified contains original features of weak information area minerals without labeled mineral types.

[0007] Geological correlation elements are extracted from the source domain mineral features and the original mineral features of the weak information area. Based on the correspondence between the source domain geological correlation elements and the weak information area geological correlation elements, a cross-domain feature mapping is constructed. The original mineral features of the weak information area are then transferred and adapted to generate intermediate mineral features of the weak information area that are adapted to the source domain feature dimensions.

[0008] Identify the sparse feature dimensions in the intermediate features of minerals in the weak information region, call the feature distribution pattern of the corresponding dimension in the source domain mineral features, and complete the sparse feature dimensions through the feature transfer mechanism of transfer learning to generate complete features of minerals in the weak information region.

[0009] The mineral completion features of the weak information area are input into a pre-trained transfer learning recognition model. The transfer learning recognition model loads the feature processing parameters obtained from source domain training and performs cross-domain recognition processing on the mineral completion features of the weak information area to generate preliminary mineral recognition results for the weak information area.

[0010] Source region samples similar to geologically related elements in the weak information area are selected from the source region mineral sample set. The features of the source region samples are input into the transfer learning recognition model to obtain the source region sample recognition results. The source region sample recognition results are compared with the labeled types of the source region samples to determine the transfer deviation. Based on the transfer deviation, the preliminary mineral recognition results of the weak information area are corrected to generate the final mineral recognition results of the weak information area. According to a second aspect of this application, a transfer learning-based intelligent mineral recognition system for weak information areas is provided. The transfer learning-based intelligent mineral recognition system for weak information areas includes a processor and a readable storage medium. The readable storage medium stores a program that, when executed by the processor, implements the aforementioned transfer learning-based intelligent mineral recognition method for weak information areas.

[0011] Based on any of the above aspects, by acquiring the source domain mineral sample set and the weak information area mineral sample set to be identified, the labeled and unlabeled mineral feature information is integrated. Then, the geological correlation elements in the mineral features of the source domain and the weak information area are extracted, and a cross-domain feature mapping is constructed. This achieves effective correlation and adaptation of features from different regions, enabling the original features of the weak information area minerals to adapt to the feature dimensions of the source domain. It identifies the sparse feature dimensions in the intermediate features of the weak information area minerals and calls the feature distribution law of the corresponding dimension in the source domain for completion. This fully utilizes the rich feature information of the source domain, solves the problem of sparse features in the weak information area, and generates more complete mineral completion features in the weak information area. The completed features are input into the pre-trained transfer learning recognition model for cross-domain recognition. The feature processing parameters obtained from the source domain training are loaded, improving the model's ability to recognize features in the weak information area. Finally, by selecting source domain samples with similar geological correlation elements, the transfer bias is determined and the preliminary recognition results are corrected, effectively reducing the recognition error caused by the difference between the source domain and the weak information area. The overall method, from data acquisition and feature processing to model recognition and result correction, significantly improves the accuracy and reliability of mineral identification in the weak information area. Attached Figure Description

[0012] Figure 1A flowchart illustrating the intelligent identification method for mineral resources in weak information areas based on transfer learning provided in an embodiment of this application is shown.

[0013] Figure 2 This paper illustrates a schematic diagram of the component structure of a mineral intelligent identification system based on transfer learning in a weak information area, as provided in an embodiment of this application. Detailed Implementation

[0014] Figure 1 The diagram illustrates a flowchart of a method for intelligent identification of mineral resources in weak information areas based on transfer learning, as provided in an embodiment of this application. It should be understood that in other embodiments, the order of some steps in the following method for intelligent identification of mineral resources in weak information areas based on transfer learning can be interchanged according to actual needs, or some steps can be omitted or deleted. The detailed steps of this method for intelligent identification of mineral resources in weak information areas based on transfer learning are described below.

[0015] Step S110: Obtain the source domain mineral sample set and the weak information area mineral sample set to be identified. The source domain mineral sample set contains source domain mineral features with labeled mineral types, and the weak information area mineral sample set to be identified contains original features of weak information area minerals without labeled mineral types.

[0016] In this embodiment, lead-zinc ore identification in the field of non-ferrous metal mineral exploration is taken as the unified application scenario. The source area mineral sample set is selected from the systematic exploration data of a known large lead-zinc ore cluster. After years of geological work, different types of lead-zinc ore bodies and associated mineralizations have been clearly identified in this area. All samples are labeled with detailed mineral type information, such as galena ore bodies, sphalerite ore bodies, and mixed lead-zinc mineralizations. The source area mineral characteristics are collected through multiple means, including stratigraphic sequence and structural trace data obtained from regional geological mapping, elemental anomaly and physical property parameter data obtained from geophysical and geochemical surveys, and core logging and ore identification data obtained from drilling projects. These data cover multiple dimensions such as stratigraphic lithology, tectonic activity, mineral assemblage, and geochemical anomalies, and have all undergone standardization and quality verification.

[0017] The sample set of minerals to be identified in the weak information area comes from a prospective lead-zinc mine survey area. Due to natural limitations, only small-to-medium scale regional geological surveys and a small amount of geophysical and geochemical exploration have been carried out in this area. The lack of systematic drilling control has resulted in significant information sparsity in the geological data. The original characteristics of minerals in the weak information area include stratigraphic unit divisions on the regional geological map, petrological identification results from a small number of outcrops, dispersed flow geochemical measurement data, and high-precision surface magnetic survey data. These data have not yet been labeled with mineral types; only preliminary analysis can identify signs of mineralization and alteration, but the specific mineral type and occurrence state cannot be determined. During data acquisition, for privacy-sensitive data such as sampling point information involving geographical location, spatial aggregation technology was used to process the data. The coordinates of the sampling points were assigned to a larger range of grid cells, so that the data only reflects the regional distribution trend and cannot pinpoint the specific sampling location, thus achieving privacy protection.

[0018] Step S120: Extract the geological correlation elements in the source domain mineral features and the geological correlation elements in the original mineral features of the weak information area. Based on the correspondence between the source domain geological correlation elements and the weak information area geological correlation elements, construct a cross-domain feature mapping. Perform migration adaptation processing on the original mineral features of the weak information area to generate intermediate mineral features of the weak information area that are adapted to the feature dimensions of the source domain.

[0019] Step S121: Extract geological correlation elements from the source region mineral characteristics. The geological correlation elements include the stratigraphic and lithological characteristics, tectonic activity characteristics, and mineral symbiotic assemblage characteristics of the region where the source region mineral is located. Each geological correlation element carries a feature identifier corresponding to the source region mineral type.

[0020] When extracting geological correlation elements from source region mineral characteristics, stratigraphic lithological characteristics specifically include the rock types (such as carbonate rocks, clastic rocks, volcanic rocks, etc.) of the stratigraphic units exposed in the source region at various ages, the rock structure (such as degree of crystallization, grain size, bedding type), the main mineral composition and content range; tectonic activity characteristics cover the morphological types of fold structures (such as upright folds, inclined folds), axial plane occurrence, hinge extension direction, mechanical properties of fault structures (such as compressive, tensile, torsional), occurrence elements, fracture zone characteristics (such as infill composition, cementation degree), development density, number of sets, and occurrence of joints and fractures; mineral association characteristics are tailored to different mineral types. For example, galena often forms associations with sphalerite, pyrite, quartz, and calcite, while sphalerite is often associated with galena, fluorite, and barite. The generation sequence, spatial distribution relationship, and embedding characteristics of minerals in each association are also included in the characteristic scope. Each extracted geological correlation element is bound to the source region mineral type through feature identification. For example, the lithological characteristics of a certain group of strata are identified as "carbonate rock-galena mineralization", indicating that the carbonate rocks and galena are directly related to the formation of galena.

[0021] Step S122: Extract geological correlation elements from the original mineral characteristics of the weak information area. The geological correlation elements include the stratigraphic lithology characteristics, tectonic activity characteristics, and mineral symbiotic assemblage characteristics of the area where the minerals in the weak information area are located. Each geological correlation element is associated with the collection scene information of the weak information area sample.

[0022] When extracting geological correlation elements from the original mineral characteristics of weak information areas, stratigraphic lithological characteristics are mainly based on the stratigraphic units of each age divided by the regional geological map, including the general types of rocks (such as limestone, sandstone, and shale) and inferred mineral composition. However, detailed lithological assemblage and thickness data of some stratigraphic units are missing or uncertain. Tectonic activity characteristics are obtained through remote sensing image interpretation and a small number of field route surveys, including the general direction and distribution of major faults in the region and the morphological signs of local folds. However, specific occurrence data (such as dip and dip angle) and mechanical property judgments are relatively rough. Mineral assemblage characteristics are based on the identification results of sporadic outcrops and a small number of artificial heavy mineral samples, which have revealed mineral grains of pyrite and galena, but systematic mineral assemblage analysis and quantitative data are lacking. Each geological correlation element in a weak information area is associated with its collection scene information. For example, the tectonic activity characteristics of a soil geochemical sample are associated with the scene description "the sampling point is located next to the linear tectonic zone interpreted by remote sensing, and the terrain slope is relatively gentle". The mineral paragenesis characteristics of a rock specimen are associated with the scene information "collected from an altered outcrop near the contact zone between carbonate rocks and clastic rocks".

[0023] Step S123: Perform cross-domain correlation analysis on the stratigraphic lithological characteristics of the source domain and the weak information area, calculate the difference in characteristic performance of the same lithological type in the source domain and the weak information area, and assign cross-domain mapping weights to lithological characteristics based on the difference values.

[0024] Step S1231: Extract all lithological types from the source region mineral characteristics and establish a source region lithological type library. Each lithological type corresponds to a set of characteristic performance data, which includes the proportion of lithological components and structural density characteristics.

[0025] When constructing the source domain lithology type database, characteristic data were collected for each lithology type identified in the source domain, such as limestone, dolomite, sandstone, shale, and granite. Lithological composition percentages include the relative content ranges of the main mineral components constituting the rock (such as calcite, dolomite, quartz, and feldspar); structural density characteristics cover physical parameters such as porosity, permeability, and bulk density. These data were obtained through laboratory testing and analysis of a large number of core samples. Each data set contains test results from multiple samples and reflects the overall characteristics of the lithology type.

[0026] Step S1232: Extract all lithological types from the original mineral characteristics of the weak information area and establish a lithological type library for the weak information area. Each lithological type also corresponds to a set of characteristic performance data, which includes the proportion of lithological components and structural density characteristics.

[0027] The establishment of a lithological type database for areas with weak information relies on limited field outcrop observations, thin section identification of rocks, and remote sensing spectral analysis. For each identified lithological type, the proportion of lithological components is estimated by the types and approximate contents of minerals identified under a microscope, while structural density characteristics are inferred based on regional geological analogies and limited physical property test data. Due to the limited sample size and testing conditions, the accuracy of some characteristic data is low, and there is a certain degree of uncertainty.

[0028] Step S1233: Perform type matching between the source domain lithology type database and the weak information area lithology type database to find lithology types with the same name or similar attributes, forming cross-domain lithology matching pairs. For each cross-domain lithology matching pair, select multiple sets of characteristic performance data of the lithology type in the source domain, calculate the mean and standard deviation of the characteristic data, and determine the standard performance range of the lithology characteristics in the source domain. Using the same method, calculate the mean and standard deviation of the characteristic performance data of the lithology type in the weak information area to determine the actual performance range of the lithology characteristics in the weak information area.

[0029] When matching lithological type databases between the source domain and the weak information region, direct name matching is performed first. For example, "limestone" in the source domain and "limestone" in the weak information region form a matching pair with the same name. For lithological types with different names but similar properties, such as "dolomite limestone" in the source domain and "dushy dolomite" in the weak information region, their mineral composition, texture, and other properties are compared to determine that they are lithological types with similar properties, forming cross-domain lithological matching pairs. For each cross-domain lithological matching pair, multiple sets of characteristic performance data for that lithological type are selected from the source domain. By calculating the mean and standard deviation of these data, the standard performance range of the lithological characteristics in the source domain is determined, i.e., the interval in which most sample data are concentrated. The same calculation method is used to process the characteristic performance data of that lithological type in the weak information region to obtain the actual performance range of the lithological characteristics in the weak information region.

[0030] Step S1234: Calculate the difference between the standard performance range and the actual performance range, wherein the difference is the ratio of the absolute difference between the two averages to the sum of their standard deviations.

[0031] For each cross-domain lithology matching pair, the difference between the standard performance range (source domain) and the actual performance range (weak information area) is calculated. Specifically, the calculation process involves first determining the absolute difference between the average value of the source domain's characteristic performance data and the average value of the weak information area's characteristic performance data; then, calculating the sum of the standard deviations of the source domain's characteristic performance data and the standard deviations of the weak information area's characteristic performance data; finally, dividing the absolute difference by the sum, the quotient being the difference value. This difference value reflects the relative degree of difference in characteristic performance between the same lithology type in the source domain and the weak information area.

[0032] Step S1235: Set a threshold for classifying the difference value, including a first threshold and a second threshold, wherein the first threshold is less than the second threshold; if the difference value is less than the first threshold, then assign a first mapping weight to the cross-domain lithology matching pair; if the difference value is greater than or equal to the first threshold and less than the second threshold, then assign a second mapping weight to the cross-domain lithology matching pair; if the difference value is greater than or equal to the second threshold, then assign a third mapping weight to the cross-domain lithology matching pair; the first mapping weight is greater than the second mapping weight, and the second mapping weight is greater than the third mapping weight.

[0033] Based on the stability and regional variability of geological characteristics, a tiered threshold for difference values ​​is set, including a first threshold and a second threshold, with the first threshold being lower than the second threshold. When the calculated difference value is lower than the first threshold, it indicates that the lithological characteristics of the source region and the weak information area are relatively similar, and a higher first mapping weight is assigned to it. When the difference value is between the first and second thresholds, it indicates a moderate difference, and a moderate second mapping weight is assigned. When the difference value is greater than or equal to the second threshold, it indicates a significant difference, and a lower third mapping weight is assigned.

[0034] Step S1236: Select lithological types in the source domain lithological type library that are similar to but not completely matched with the lithological type attributes of the weak information area, calculate their characteristic difference values ​​with the lithological types of the weak information area, and assign auxiliary mapping weights according to the correspondence between the above difference value classification threshold and mapping weights. The value of the auxiliary mapping weight is lower than the first mapping weight, second mapping weight or third mapping weight of the corresponding difference value interval.

[0035] For certain lithological types in the weak information region that cannot be directly matched with the source domain lithological types or have low attribute similarity, lithological types with similar attributes are selected from the source domain lithological type database as auxiliary references. The characteristic difference values ​​between these auxiliary lithological types and the lithological types in the weak information region are calculated. Auxiliary mapping weights are also assigned according to the correspondence between the difference value classification threshold and the mapping weight, but the values ​​of the auxiliary mapping weights are lower than the main mapping weights (i.e., the first, second, or third mapping weights) of the corresponding difference value intervals, so as to reflect their status as auxiliary references.

[0036] Step S1237: The initial cross-domain mapping weight and the auxiliary mapping weight are weighted and fused. During the fusion process, the lithology type of the weak information area is referenced in the collection scene information. If the similarity between the collection scene and the source domain meets the preset similarity requirements, the proportion of the initial mapping weight in the fusion is increased.

[0037] The initial cross-domain mapping weights obtained through direct matching are weighted and fused with the auxiliary mapping weights obtained through assisted matching. During the fusion process, the data collection scenario information of lithology types in the weak information area is taken into account, such as the geological background of the sampling points (e.g., sedimentary environment, tectonic location), topographic conditions, etc. If the data collection scenario of a certain lithology type in the weak information area has a high similarity to the formation environment of the corresponding lithology type in the source domain (meeting the preset similarity requirements), the proportion of the initial cross-domain mapping weight in the fusion is increased; otherwise, the proportion of the auxiliary mapping weight is appropriately increased to enhance the rationality of the weights after fusion.

[0038] Step S1238: Based on the fused weight values, determine the cross-domain mapping weight of lithological characteristics for each cross-domain lithology matching pair. The sum of the weight values ​​is set to a fixed total weight value to ensure a balanced weight distribution among different lithology types.

[0039] The weight values ​​after fusion are normalized, and the sum of the weight values ​​of all cross-domain lithology matching pairs is set to a fixed total weight value (e.g., unit 1) to ensure a balanced weight distribution for different lithology types and to avoid the weight of a certain lithology type being too high or too low, which would affect the effect of subsequent feature transfer.

[0040] Step S1239: Record the feature difference values ​​and weight allocation process of each cross-domain lithology matching pair to form a cross-domain mapping weight table of lithology features.

[0041] The detailed information of each cross-domain lithology matching pair is recorded in a structured manner, including lithology type name, the calculation process of feature difference value (mean, standard deviation, difference value), initial mapping weight, auxiliary mapping weight, weight ratio during fusion, weight value after fusion, final weight after normalization, etc., forming a cross-domain mapping weight table of lithology features, which can be queried and called in subsequent migration and adaptation processing of stratigraphic lithology features in weak information areas.

[0042] Step S124: Using the same method, calculate the cross-domain performance differences of tectonic activity characteristics and mineral symbiotic assemblage characteristics between the source domain and the weak information region, and assign cross-domain mapping weights to the corresponding characteristics.

[0043] For tectonic activity characteristics, taking "fracture structures" as an example, the characteristic data of fracture structures in the source domain include fracture length, strike stability, dip angle variation, breccia width, infill composition, and spatial relationship with mineralization. The characteristic data of fracture structures in areas with weak information are based on remote sensing interpretation and limited field surveys, including the approximate extension direction of the fracture, linear image clarity, and inferred breccia extent. Using the same processing procedure as for stratigraphic and lithological characteristics, including establishing a tectonic activity characteristic type database, matching cross-domain tectonic activity characteristic pairs, calculating characteristic difference values, assigning initial and auxiliary mapping weights, fusing weights and normalizing them, the cross-domain mapping weights of tectonic activity characteristics are finally obtained.

[0044] The processing of mineral paragenesis characteristics is similar. Taking the "galena-pyrite paragenesis" as an example, the characteristic data of this assemblage in the source domain include the content ratio of the two minerals, their grain size relationship, their embedding patterns (such as inclusions and adjacency), and their relationship with gangue minerals. Similar paragenesis assemblages were obtained in the weak information region through identification of a small number of samples, but the data is relatively scattered and lacks systematicity. The difference value was calculated and cross-domain mapping weights were assigned according to the above process to ensure that tectonic activity characteristics and mineral paragenesis characteristics also obtained appropriate migration weights.

[0045] Step S125: Integrate the cross-domain mapping weights of stratigraphic lithological characteristics, tectonic activity characteristics, and mineral symbiotic assemblage characteristics to construct a cross-domain feature mapping matrix between the source domain and the weak information area. Each element in the cross-domain feature mapping matrix corresponds to a set of mapping relationships and weights of geological correlation elements between the source domain and the weak information area.

[0046] The cross-domain mapping weights calculated separately for stratigraphic lithology, tectonic activity, and mineral assemblage characteristics are integrated to construct a three-dimensional cross-domain feature mapping matrix. The row dimensions of the matrix correspond to the types of geological associated elements in the source domain (subdivided into specific feature items, such as limestone, NE-trending faults, galena-pyrite assemblage, etc.), while the column dimensions correspond to the types of geological associated elements in the weak information area (also subdivided into specific feature items). Each element in the matrix represents the cross-domain mapping weight between a certain geological associated element in the source domain and a certain geological associated element in the weak information area. This matrix clearly shows the mapping relationships and weights between various types of geological associated elements in the source domain and the weak information area.

[0047] Step S126: Input the original mineral features of the weak information area into the cross-domain feature mapping matrix, and adjust the feature values ​​of each geological associated element in the weak information area according to the corresponding mapping weights so that the adjusted feature values ​​conform to the representation rules of the source domain features.

[0048] Each geologically related element's feature value (such as the thickness of a stratigraphic unit, the strike angle of a fault, the percentage content of a mineral, etc.) in the original mineral characteristics of the weak information area is multiplied one by one with the corresponding mapping weight in the cross-domain feature mapping matrix to achieve initial migration adjustment. During the adjustment process, the feature representation rules of the corresponding geologically related elements in the source domain (such as the numerical range and trend of such elements in the source domain) are also referenced to verify the adjusted values. If the adjusted values ​​exceed the reasonable representation range of the source domain features, a second correction is performed to make them fall within the numerical range that the source domain features usually exhibit, ensuring that the adjusted feature values ​​conform to the feature representation rules of the source domain.

[0049] Step S127: Perform migration adaptation verification on the adjusted weak information area features, calculate the similarity between the adjusted weak information area geological association feature features and the source domain similar geological association feature features, and compare the similarity with the preset migration adaptation threshold.

[0050] Using appropriate similarity calculation methods (such as cosine similarity, correlation coefficient, etc.), the similarity between the adjusted feature vectors of geologically related elements in the weak information area and the feature vectors of similar geologically related elements in the source domain is calculated. A migration adaptation threshold is preset, which is comprehensively set based on the dispersion of the source domain data and the importance of the features. The calculated similarity value is compared with the preset migration adaptation threshold to determine whether the adjusted features of the weak information area have met the adaptation requirements with the features of the source domain.

[0051] Step S128: If the similarity is lower than the preset migration adaptation threshold, the cross-domain mapping weight of the corresponding geological associated elements is readjusted, and the feature values ​​of the geological associated elements in the weak information area are migrated and adjusted again.

[0052] When the similarity between the adjusted features of a geologically related element in a weak information area and similar features in the source domain is lower than a preset migration adaptation threshold, it indicates that the migration adaptation effect of that element is poor. In this case, the process returns to the cross-domain mapping weight calculation stage for that geologically related element, re-evaluating the differences in its feature performance, checking the accuracy of the difference calculation and the rationality of the weight allocation (e.g., whether the proportion of auxiliary mapping weights needs adjustment), and thus obtaining new cross-domain mapping weights. Using these new weights, the feature values ​​of that geologically related element in the weak information area are adjusted again, and the similarity is recalculated until the similarity reaches the preset threshold.

[0053] Step S129: If the similarity reaches the preset migration adaptation threshold, cross-domain collaborative integration is carried out on all adjusted geological associated element features. By adjusting the numerical distribution of geological associated element features in the weak information area, the feature association relationship between different geological associated elements conforms to the association relationship of the source domain features.

[0054] For example, step S1291: Extract the correlation data between stratigraphic lithology characteristics, tectonic activity characteristics, and mineral symbiotic assemblage characteristics from the source region mineral characteristics, calculate the correlation coefficient between stratigraphic lithology characteristics, tectonic activity characteristics, and mineral symbiotic assemblage characteristics, and establish a source region characteristic correlation matrix, where the matrix elements represent the correlation strength between the two types of characteristics.

[0055] From the source region mineral characteristics, based on a large sample of data, multiple characteristic variables were extracted for each of the stratigraphic lithology, tectonic activity, and mineral assemblage characteristics. Using statistical analysis methods (such as the Pearson correlation coefficient method), the correlation coefficients between each characteristic variable in each pair of characteristics (stratigraphic lithology and tectonic activity, stratigraphic lithology and mineral assemblage, tectonic activity and mineral assemblage) were calculated. These correlation coefficients reflect the strength and direction of the linear association between different characteristic variables. The correlation coefficients were arranged in matrix form to establish a source region characteristic correlation matrix, where the rows and columns correspond to different categories of characteristic variables, and the matrix element values ​​represent the correlation strength (correlation coefficient value) between the corresponding characteristic variables.

[0056] Step S1292: Extract three types of geological correlation features from the adjusted weak information area features, calculate the correlation coefficients between the three types of geological correlation features in the weak information area features, and establish the initial correlation matrix of the weak information area.

[0057] From the features of the weak information area after migration adaptation, feature variables corresponding to stratigraphic lithology, tectonic activity, and mineral assemblage characteristics in the source domain feature correlation matrix are extracted. Using the same correlation coefficient calculation method as the source domain, the correlation coefficients between these three types of geological correlation element feature variables in the weak information area are calculated, thereby establishing an initial correlation matrix for the weak information area. The structure of this matrix (the feature variables corresponding to rows and columns) is completely consistent with the source domain feature correlation matrix for comparative analysis.

[0058] Step S1293: Calculate the matrix difference between the source domain feature correlation matrix and the initial correlation matrix of the weak information region, wherein the matrix difference is the square root of the sum of squares of the differences of the corresponding elements.

[0059] Let A be the source domain feature correlation matrix and B be the initial correlation matrix of the weak information region, both having the same dimension. To calculate the matrix difference, first calculate the difference between corresponding elements in A and B (A[i][j] - B[i][j]). Then, square each difference to obtain a square matrix. Sum the values ​​of all elements in this square matrix to obtain a sum of squares. Finally, take the square root of this sum of squares, which is the matrix difference between the source domain feature correlation matrix and the initial correlation matrix of the weak information region. This matrix difference comprehensively reflects the overall degree of difference between the two matrices.

[0060] Step S1294: Set an association consistency threshold and compare the matrix difference value with the association consistency threshold. If the matrix difference value is lower than the association consistency threshold, it means that the association relationship of the weak information region features is consistent with the association relationship of the source domain features, and no collaborative integration is required.

[0061] A correlation consistency threshold is set, which is determined based on the stability of the source domain feature correlation matrix and the quality of the data in the weak information area. The calculated matrix difference value is compared with the correlation consistency threshold. If the matrix difference value is lower than the correlation consistency threshold, it indicates that the correlation between the three types of geological correlation elements in the weak information area is already close to the correlation between the source domain features, which meets the requirements of collaborative integration, and no further adjustment is needed.

[0062] Step S1295: If the matrix difference value is higher than the correlation consistency threshold, calculate the difference between each corresponding element in the source domain feature correlation matrix and the initial correlation matrix of the weak information region. Matrix elements with element differences greater than the preset element difference threshold are identified as matrix elements that need to be adjusted. Analyze the feature correlation pairs corresponding to the matrix elements that need to be adjusted to determine the feature correlation relationships of the weak information region that need to be adjusted.

[0063] When the matrix difference value exceeds the correlation consistency threshold, it indicates that the correlation relationship between the features of the weak information area and the features of the source domain still differs significantly. In this case, the difference between corresponding elements of the source domain feature correlation matrix and the initial correlation matrix of the weak information area is calculated. A preset element difference threshold is set, and matrix elements with differences greater than this threshold are marked as elements requiring adjustment. Each matrix element requiring adjustment corresponds to a pair of feature variables (from different categories of geological correlation elements), i.e., a feature correlation pair. By analyzing these feature correlation pairs, it is determined which feature variables in the weak information area require adjustment in their correlation relationships.

[0064] Step S1296: For the feature association pairs that need to be adjusted, extract the association strength parameter of the feature association pair in the source domain, and calculate the target association strength of the feature association pair in the weak information region based on the association strength parameter.

[0065] For each feature association pair that needs adjustment (such as a lithological feature variable of a certain stratigraphy and a tectonic activity feature variable), the correlation strength parameter (correlation coefficient value) corresponding to the feature association pair is extracted from the source domain feature correlation matrix. Based on the source domain correlation strength parameter and combined with the current correlation strength of the weak information area, the target correlation strength that the feature association pair in the weak information area needs to achieve is calculated. The target correlation strength should be as close as possible to the correlation strength of the source domain, while taking into account the actual situation of the data in the weak information area.

[0066] Step S1297: Based on the target correlation strength, adjust the numerical distribution of the two types of geological correlation feature characteristics corresponding to the weak information area. During the adjustment process, keep the numerical range of a single feature dimension in line with the source domain adaptation requirements. After adjustment, recalculate the correlation matrix of the weak information area features and compare the new matrix difference value with the correlation consistency threshold.

[0067] Based on the calculated target correlation strength, the numerical distribution of the two types of geological correlation elements constituting the feature correlation pair in the weak information area is adjusted. This adjustment can be achieved by appropriately scaling, shifting, or transforming the numerical values ​​of the feature variables, but it is necessary to ensure that the numerical range of a single feature dimension remains within the range required by the source domain after the previous migration and adaptation process, avoiding disruption of the existing dimensional adaptation effect. After adjustment, the correlation coefficients between the three types of geological correlation elements in the weak information area are recalculated, a new weak information area correlation matrix is ​​established, and the new matrix difference value is calculated and compared with the correlation consistency threshold.

[0068] Step S1298: If the new matrix difference value is still higher than the association consistency threshold, continue to calculate the difference of each corresponding element in the new association matrix, find the matrix elements whose element difference is greater than the preset element difference threshold, and adjust their corresponding feature association pairs until the matrix difference value is lower than the association consistency threshold.

[0069] If the adjusted matrix difference value is still higher than the association consistency threshold, repeat the above process: calculate the difference between corresponding elements of the new association matrix and the source domain association matrix, identify the elements that need adjustment and their corresponding feature association pairs, calculate the target association strength, adjust the feature value distribution, and recalculate the association matrix and matrix difference value. Iterate in this way until the matrix difference value is reduced below the association consistency threshold.

[0070] Step S1299: After completing the correlation adjustment, perform overall collaborative verification on all geological correlation feature characteristics in the weak information area, calculate the overall similarity between the adjusted features and the source domain features in the overall correlation structure, compare the overall similarity with the preset standard, and if the overall similarity meets the preset standard, the cross-domain collaborative integration is completed; if the overall similarity does not meet the preset standard, re-examine the adjustment process of the feature correlation matrix, correct the adjustment parameters, and integrate again.

[0071] After the feature association relationship adjustment is completed and the matrix difference value is below the association consistency threshold, a comprehensive collaborative verification is performed on all geological association element features in the weak information area. Appropriate methods (such as structural similarity based on principal component analysis, network structure similarity based on graph theory, etc.) are used to calculate the overall similarity between the adjusted weak information area features and the source domain features in terms of overall association structure. This considers not only pairwise feature associations but also the overall association pattern formed by all features. A pre-defined overall similarity standard is set. If the calculated overall similarity meets this standard, the cross-domain collaborative integration is considered complete; otherwise, the previous feature association matrix adjustment process needs to be re-examined to analyze whether there are any improperly adjusted or omitted feature association pairs. After correcting the adjustment parameters (such as the calculation method of the target association strength, the magnitude of numerical adjustment, etc.), the collaborative integration operation is performed again.

[0072] Step S1210: Based on the collaboratively integrated features, generate intermediate features of mineral resources in weak information areas that are adapted to the feature dimensions of the source domain. Each geological correlation element of the intermediate features of mineral resources in weak information areas carries a weight identifier for source domain migration adaptation.

[0073] After completing cross-domain collaborative integration and passing overall collaborative verification, the features of all geologically related elements (stratigraphic lithology, tectonic activity, and mineral assemblages) are combined after migration adaptation and collaborative integration to form intermediate mineral features for weak information areas. These intermediate mineral features for weak information areas are completely consistent with the source domain mineral features in terms of feature dimensions (including the number, type, and order of feature variables), ensuring that they can be directly input into models trained based on the source domain for subsequent processing. Simultaneously, each geologically related element carries a source domain migration adaptation weight identifier, which records the source information of the cross-domain mapping weights used by the element during the migration adaptation process (such as which cross-domain matching pair it comes from, the fusion weight value, etc.), facilitating subsequent traceability and analysis.

[0074] Step S130: Identify the sparse feature dimensions in the intermediate features of the minerals in the weak information region, call the feature distribution pattern of the corresponding dimension in the mineral features of the source domain, and complete the sparse feature dimensions through the feature transfer mechanism of transfer learning to generate the complete features of the minerals in the weak information region.

[0075] Step S131: Count the number of effective features in each feature dimension of the intermediate features of minerals in the weak information area, calculate the proportion of effective features, and mark the feature dimensions with an effective feature proportion lower than the preset sparsity threshold as sparse feature dimensions to form a list of sparse feature dimensions.

[0076] For each feature dimension (i.e., each specific feature variable, such as the porosity of a certain type of rock, the dip angle of a certain fracture, the content of a certain mineral, etc.) of the intermediate features of minerals in weak information areas, the number of effective features is statistically analyzed. The number of effective features refers to the number of samples with valid observations or reasonable inferences under that dimension, excluding missing values ​​or outliers. The ratio of the number of effective features to the total number of samples in that dimension is calculated, i.e., the proportion of effective features. A sparse threshold is preset, which is set according to the importance of the features and the difficulty of data collection. Feature dimensions with an effective feature proportion lower than the preset sparse threshold are marked as sparse feature dimensions, and all sparse feature dimensions are arranged in order of their position in the feature vector to form a list of sparse feature dimensions.

[0077] Step S132: Extract source domain feature data corresponding to each dimension in the sparse feature dimension list from the source domain mineral features, and analyze the numerical distribution pattern of the source domain feature data. The numerical distribution pattern includes the concentration range of feature values, distribution density, and correlation relationship between adjacent dimensions.

[0078] Based on the list of sparse feature dimensions, all source domain feature data corresponding to each sparse feature dimension are extracted from the source domain mineral features. In-depth statistical analysis is then performed on the extracted source domain feature data to reveal its numerical distribution patterns. Specifically, this includes: the concentration range of feature values, i.e., the range of values ​​where most data points are concentrated; the distribution density, i.e., the density of data points within different numerical ranges, which can be described by a probability density function or histogram; and the correlation between this feature dimension and adjacent dimensions (referring to other feature dimensions that are close in position in the feature vector or geologically related), such as positive correlation, negative correlation, or no obvious correlation, and the strength of the correlation.

[0079] Step S133: For each sparse feature dimension, construct a transfer completion model based on the feature distribution pattern of the corresponding dimension in the source domain. The transfer completion model loads the feature distribution parameters of the source domain and has the ability to transfer the distribution pattern of the source domain to the sparse dimension of the weak information region.

[0080] For each sparse feature dimension, a specialized transfer completion model is constructed using the feature distribution patterns obtained from the source domain analysis. The core of this transfer completion model is loading source domain feature distribution parameters. These parameters characterize the numerical distribution patterns of the source domain, such as boundary values ​​of the concentration intervals, parameters of the distribution density function (e.g., mean and variance for a normal distribution), and correlation coefficients with adjacent dimensions. Through these parameters, the transfer completion model can transfer the distribution patterns of the source domain to the sparse feature dimensions of the weakly informationd region, thereby guiding the prediction and completion of missing data.

[0081] Step S134: Input the existing effective feature data in the sparse feature dimension of the weak information region into the transfer completion model. The transfer completion model performs distribution adaptation on the effective feature data based on the distribution law of the source domain to determine the feature distribution trend of the sparse dimension of the weak information region.

[0082] The existing valid feature data in a sparse feature dimension of a weak information region is input into the transfer completion model it constructs. The model first adapts the distribution of the valid feature data in the weak information region based on the loaded source domain feature distribution pattern, making the valid data in the weak information region conform to the distribution pattern of the source domain as closely as possible (e.g., adjusting the mean and variance of the data to approach the source domain, but not completely identical to it). Through this distribution adaptation, the model can analyze and determine the overall feature distribution trend of that sparse feature dimension in the weak information region, including the approximate direction of data concentration, possible peak positions, and the approximate pattern of numerical changes with samples.

[0083] Step S135: Based on the feature distribution trend, perform numerical prediction on the location of missing data in the sparse feature dimension of the weak information region. During the prediction process, refer to the correlation relationship between adjacent dimensions in the source domain features to ensure that the predicted value is consistent with the existing effective feature data in the weak information region.

[0084] Based on the characteristic distribution trend of the sparse feature dimension in the identified weak information region, the model numerically predicts the location of all missing data in that dimension. During the prediction process, the model considers not only the distribution trend of the current sparse dimension itself but also the correlation between that dimension and adjacent dimensions in the source domain features. For example, if the sparse dimension in the source domain is significantly positively correlated with a certain adjacent dimension, then when predicting in the weak information region, if the adjacent dimension has a high value, the predicted value of the sparse dimension should also tend to be high. This ensures that the predicted value maintains consistency with the existing valid feature data in the weak information region in terms of correlation, avoiding isolated predictions that do not conform to geological logic.

[0085] Step S136: Fill the missing data positions with the predicted values ​​to complete the preliminary completion of the sparse feature dimensions and obtain the preliminary completed features.

[0086] The values ​​predicted by the transfer completion model are then filled into the corresponding missing data positions in the sparse feature dimensions of the weak information region. This completion operation is performed on all feature dimensions marked as sparse. After completion, the sparse feature dimensions in the intermediate features of minerals in the weak information region no longer have missing values, resulting in preliminary feature completion.

[0087] Step S137: Perform a migration consistency check on the preliminary completed features, calculate the matching degree between the completed weak information region features and the corresponding dimension features of the source domain in terms of distribution patterns, and compare the matching degree with the preset consistency threshold.

[0088] Step S1371: Extract source domain feature data corresponding to sparse feature dimensions from source domain mineral features, analyze the distribution pattern of source domain feature data using statistical methods, and generate source domain feature distribution curves. The source domain feature distribution curves include parameters such as numerical frequency, peak position, and distribution width.

[0089] Source domain feature data corresponding to each dimension in the sparse feature dimension list is extracted from the source domain mineral characteristics. Appropriate statistical methods (such as histogram methods, kernel density estimation methods, etc.) are used to analyze the source domain feature data, generating a source domain feature distribution curve. This curve visually displays the distribution pattern of the source domain feature data, including key information such as numerical frequency (the number of times or probability of data points appearing within different numerical intervals), peak position (the value corresponding to the highest point of the curve, i.e., the position where the data is most concentrated), and distribution width parameter (the numerical range of the curve from the left start point to the right end point, or a parameter reflecting the degree of data dispersion).

[0090] Step S1372: Extract the corresponding sparse dimension completion feature data from the initial completion features, and use the same statistical method to generate the completion feature distribution curve, so that the statistical dimensions of the two distribution curves are consistent.

[0091] Extract the completed feature data corresponding to the aforementioned sparse dimensions from the initial completed features. Analyze the completed feature data using the same statistical methods (including the same interval division method, kernel function type, and parameter settings) as the source domain feature distribution curve to generate a completed feature distribution curve. Ensure that the statistical dimensions (such as the numerical range of the horizontal axis and interval intervals) of the two distribution curves (source domain and completed) are completely consistent for accurate comparison.

[0092] Step S1373: Calculate the similarity between the source domain feature distribution curve and the completed feature distribution curve. The similarity is calculated by comprehensively considering the percentage of overlapping area of ​​the curves, the peak position deviation, and the distribution width deviation. The result is the matching degree.

[0093] The similarity between the source domain feature distribution curve and the completed feature distribution curve is calculated. This similarity considers several aspects: the percentage of overlapping area (the proportion of the area of ​​the overlapping part of the two curves to the sum of the areas of the two curves), the peak position deviation (the distance between the peak position of the completed curve and the peak position of the source domain curve), and the distribution width deviation (the difference between the distribution width of the completed curve and the distribution width of the source domain curve). These three indicators are calculated using a weighted method, and the result is the matching degree. The higher the matching degree, the more similar the completed feature distribution is to the source domain feature distribution.

[0094] Step S1374: Compare the matching degree with the preset consistency threshold. If the matching degree meets the preset consistency threshold requirement, the preliminary completed feature is determined to have passed the migration consistency check. If the matching degree does not meet the preset consistency threshold requirement, analyze the difference position of the two distribution curves, calculate the average difference between the completed feature value and the source domain feature value in each numerical interval of the completed feature distribution curve, and determine the interval with the average difference higher than the preset difference threshold as the numerical interval that needs to be adjusted.

[0095] A consistency threshold is preset, set based on the distribution stability of the source domain data and the required completion accuracy. The calculated matching degree is compared with the preset consistency threshold. If the matching degree is greater than or equal to the threshold, it is determined that the initial completed features have reached the consistency requirement with the source domain features in terms of distribution patterns, and the transfer consistency check is passed. If the matching degree is less than the threshold, it is necessary to analyze the differences between the two distribution curves and identify areas where the completed curve deviates significantly from the source domain curve. Specifically, the distribution curve is divided into several numerical intervals, and the difference between the average value of the completed feature values ​​and the average value of the source domain feature values ​​in each interval (average difference) is calculated. A preset difference threshold is set, and the intervals with average differences higher than this threshold are identified as numerical intervals that require focused adjustment.

[0096] Step S1375: For the numerical range that needs to be adjusted, adjust the weight of the corresponding source domain parameter in the transfer completion model, reduce the influence weight of the source domain parameter in the numerical range, and increase the weight of the source domain parameter that matches the existing effective feature data in the weak information area. The adjustment range is based on the ratio of the average difference of the numerical range to the preset difference threshold. The higher the ratio, the greater the adjustment range.

[0097] For numerical intervals requiring focused adjustment, adjust the weights of the corresponding source domain distribution pattern parameters in the transfer completion model. For example, if the completion curve's value in a certain interval is significantly higher than the source domain curve, it indicates that the source domain parameters (such as the distribution density of that interval) have too much influence on the completion result, and their weight should be reduced. Simultaneously, increase the weights of source domain parameters that match the existing effective feature data in the weak information region (such as the distribution parameters of source domain subsamples more similar to the distribution of effective data in the weak information region). The adjustment magnitude is determined based on the ratio of the average difference of the numerical interval to a preset difference threshold. A higher ratio indicates a greater difference, and the adjustment magnitude should be larger to more effectively correct the distribution pattern of that interval.

[0098] Step S1376: After adjusting the parameter weights, the effective feature data of the sparse feature dimension of the weak information region is re-input into the transfer completion model. The transfer completion model re-predicts the missing data position based on the adjusted parameter weights, generates a distribution curve for the re-predicted completed feature data, and calculates the matching degree between the distribution curve and the source domain feature distribution curve.

[0099] After adjusting the parameter weights of the transfer completion model, the original valid feature data from the sparse feature dimension of the weak information region are re-input into the model. Based on the adjusted parameter weights, the model performs another numerical prediction of the missing data locations. Using the re-predicted completed feature data, the completed feature distribution curve is regenerated according to the previous statistical method, and the matching degree between the new curve and the source domain feature distribution curve is calculated.

[0100] Step S1377: Repeat the steps of adjusting the weights of the corresponding source domain parameters in the migration completion model, re-inputting the effective feature data into the migration completion model for numerical prediction, and calculating the matching degree between the new distribution curve and the source domain feature distribution curve, until the matching degree of the completed features meets the preset consistency threshold requirement.

[0101] The process involves repeatedly adjusting model parameter weights, predicting missing data values, generating feature distribution curves for completion, and calculating the matching degree. After each adjustment, the matching degree is checked to see if it has reached a preset consistency threshold. If not, the differences are analyzed and parameters are adjusted further; if it has, the iteration stops. Throughout the process, detailed records are kept of the specific magnitude of each parameter adjustment (e.g., the amount of change in a source domain parameter weight from its original value to a new value), the matching degree values ​​before and after the adjustment, and the trend of the matching degree change. This information forms a migration completion model parameter adjustment log, used to trace the completion process and model optimization.

[0102] Step S138: If the matching degree reaches the preset consistency threshold, cross-dimensional collaborative verification is performed on the completion results of all sparse feature dimensions, and the correlation between different sparse dimensions after completion is analyzed so that the correlation conforms to the correlation rules of the source domain features.

[0103] Once the matching degree of the completed features reaches a preset consistency threshold, after passing the transfer consistency check, cross-dimensional collaborative verification is performed on the completion results of all sparse feature dimensions. The analysis examines whether there are appropriate correlations between the completed sparse feature dimensions. These correlations should refer to the correlation patterns between corresponding dimensions in the source domain features. For example, if there is a negative correlation between two sparse dimensions (let's say dimension A and dimension B) in the source domain, then these two dimensions in the completed weak information region should also show a similar negative correlation trend. By calculating the correlation coefficients and partial correlation coefficients between the completed sparse dimensions, the correlation is checked to see if it conforms to the source domain patterns. If there are discrepancies, the completion results of the relevant dimensions need to be fine-tuned until the correlations between all sparse dimensions are consistent with the correlation patterns of the source domain features.

[0104] Step S139: After collaborative verification, the completed sparse feature dimensions are integrated with the original non-sparse feature dimensions to generate mineral completion features for weak information regions. Each dimension of the mineral completion features for weak information regions carries a migration identifier of the source domain distribution pattern.

[0105] After all the completion results for sparse feature dimensions have passed cross-dimensional collaborative verification, the completed sparse feature dimensions are integrated with the original non-sparse feature dimensions (i.e., feature dimensions whose effective feature ratio is higher than the preset sparsity threshold and have not undergone completion processing) to form a complete feature vector. The generated mineral completion features for weak information areas have no missing values ​​in each feature dimension, and each dimension carries a migration identifier of the source domain distribution pattern. This migration identifier records whether the dimension has undergone completion processing, the source domain distribution pattern parameters referenced during the completion process, the adjustment log number of the completion model, and other information for use in subsequent model processing and result analysis.

[0106] Step S140: Input the mineral completion features of the weak information area into the pre-trained transfer learning recognition model. The transfer learning recognition model loads the feature processing parameters obtained from the source domain training and performs cross-domain recognition processing on the mineral completion features of the weak information area to generate preliminary mineral recognition results for the weak information area.

[0107] Step S141: Input the mineral completion features of the weak information area into the feature transfer layer of the transfer learning recognition model. The feature transfer layer calls the linear transformation parameters and activation function parameters obtained from the source domain training to perform linear transformation and nonlinear activation operations on the weak information area completion features to generate the transformed features.

[0108] The mineral feature completion features in weak information areas are first input into the feature transfer layer of the transfer learning recognition model. The feature transfer layer is the first processing unit of the model, and its core function is to map the input features from the original space to a feature space more conducive to subsequent recognition. This feature transfer layer calls the linear transformation parameters learned and saved during the source domain training, including the weight matrix and bias vector. A linear transformation of the input features is achieved by performing matrix multiplication between the completed feature vector and the weight matrix, and adding the bias vector. Subsequently, activation function parameters (such as ReLU function, LeakyReLU function, etc.) determined during source domain training are applied to the linear transformation result for nonlinear activation, introducing nonlinear factors to enhance the model's ability to express complex feature relationships and generate the transformed features.

[0109] Step S142: Input the transformed features into the cross-domain processing layer. The cross-domain processing layer calculates the difference measure between the transformed features and the source domain training feature set extracted from the model storage in terms of statistical distribution. Based on the difference measure, the transformed features are numerically adjusted through a trainable adaptation network to generate domain-adapted features.

[0110] Step S1421: Obtain the transformed weak information region features from the feature transfer layer of the transfer learning recognition model, and at the same time extract the source domain training feature sample set stored in the model, select representative samples from the source domain training feature sample set to form the source domain feature reference set.

[0111] The cross-domain processing layer receives the transformed weak-information region features from the feature transfer layer. Simultaneously, it extracts the source domain training feature sample set used during the source domain training phase from the model's storage module. To improve computational efficiency and ensure representativeness, representative samples are selected from the source domain training feature sample set. Selection methods can include random sampling (ensuring sufficient sample size), stratified sampling (sampling according to mineral type proportions), or cluster center sampling (selecting cluster center samples for each category), forming the source domain feature reference set.

[0112] Step S1422: Standardize the source domain feature reference set and the transformed weak information region features to eliminate the dimensional differences between different feature dimensions.

[0113] Since the features of the source domain and the weak information region may originate from different measurement methods or have different physical meanings, the dimensions of their feature dimensions may differ. To fairly compare the differences in feature distribution, the source domain feature reference set and the transformed weak information region features are standardized. Standardization typically employs the Z-score standardization method, which involves subtracting the mean of that dimension from the source domain feature reference set, and then dividing by the standard deviation of that dimension, resulting in a mean of 0 and a standard deviation of 1 for the processed data, thus eliminating the dimensional differences between different feature dimensions.

[0114] Step S1423: Calculate the feature distance between the standardized weak information region features and each sample in the standardized source domain feature reference set, and obtain a set of feature distance values ​​using the distance calculation method.

[0115] Using appropriate distance calculation methods (such as Euclidean distance, Manhattan distance, cosine distance, etc.), the feature distance between the standardized weak information region feature vector and the feature vector of each sample in the standardized source domain feature reference set is calculated. For a weak information region feature sample, the distance values ​​with each sample in the source domain feature reference set are obtained, forming a set of feature distance values. These distance values ​​reflect the similarity between the weak information region feature and each source domain reference sample.

[0116] Step S1424: Calculate the domain difference value based on the feature distance value, where the domain difference value is the average of all feature distance values.

[0117] To comprehensively measure the difference between the overall distribution of features in the weak information region and the source domain features, the above set of feature distance values ​​are averaged. The average value is the domain difference value. The smaller the domain difference value, the closer the overall distribution of features in the weak information region is to the source domain feature reference set; conversely, the larger the value, the greater the difference.

[0118] Step S1425: Set the domain difference value classification standard, including a first-level threshold, a second-level threshold, and a third-level threshold. The first-level threshold is less than the second-level threshold, and the second-level threshold is less than the third-level threshold. If the domain difference value is less than the first-level threshold, a first-level adaptation adjustment strategy is applied, with the adjustment range being the minimum. If the domain difference value is greater than or equal to the first-level threshold and less than the second-level threshold, a second-level adaptation adjustment strategy is applied, with the adjustment range being the medium. If the domain difference value is greater than or equal to the second-level threshold and less than the third-level threshold, a third-level adaptation adjustment strategy is applied, with the adjustment range being the maximum.

[0119] Based on the magnitude of the domain difference value, a three-level classification standard is set: a first-level threshold, a second-level threshold, and a third-level threshold, with the first-level threshold < the second-level threshold < the third-level threshold. Different domain difference value levels correspond to different domain adaptation adjustment strategies: when the domain difference value is less than the first-level threshold, it indicates that the cross-domain difference is very small, and the first-level domain adaptation adjustment strategy is adopted, with the smallest adjustment range, mainly performing fine distribution calibration; when the domain difference value is between the first-level and second-level thresholds, the second-level domain adaptation adjustment strategy is adopted, with a moderate adjustment range, performing a more significant distribution adjustment; when the domain difference value is between the second-level and third-level thresholds, it indicates that the cross-domain difference is large, and the third-level domain adaptation adjustment strategy is adopted, with the largest adjustment range, performing a significant distribution transformation.

[0120] Step S1426: Based on the level of the domain difference value of the transformed weak information region features, call the corresponding domain adaptation adjustment strategy. The domain adaptation adjustment strategy includes the adjustment coefficient of the feature dimension, and the adjustment coefficient is determined based on the dimensional importance of the source domain training features.

[0121] Based on the domain difference value calculated from the features of the weak information region after transformation, its level is determined, and then the corresponding domain adaptation adjustment strategy is applied. Each domain adaptation adjustment strategy contains a set of adjustment coefficients for each feature dimension. These adjustment coefficients are determined during the source domain training process by analyzing the contribution of different feature dimensions to mineral type identification (e.g., based on feature importance assessment methods). For important feature dimensions in the source domain that have a significant impact on the identification results, larger adjustment coefficients are assigned to focus on adjusting these dimensions during the adaptation process, ensuring that their distribution is as close as possible to the source domain.

[0122] Step S1427: Adjust the values ​​of each dimension of the transformed weak information region features according to the adjustment coefficient, while keeping the correlation between feature dimensions unchanged during the adjustment process.

[0123] Based on the feature dimension adjustment coefficients in the invoked domain adaptation adjustment strategy, the values ​​of each dimension of the transformed weak information region features are adapted and adjusted. The adjustment method can be multiplying the feature value by the adjustment coefficient, or adding / subtracting the offset determined by the adjustment coefficient. During the adjustment process, a collaborative adjustment of all relevant dimensions is adopted to ensure that the original relationships between feature dimensions (such as those established through cross-domain collaborative integration) are not disrupted, thus maintaining the overall structure of the features.

[0124] Step S1428: Perform domain offset verification on the adjusted features, standardize the adjusted features using the same standardization parameters as the source domain feature reference set, and then calculate the domain difference value between the standardized adjusted features and the standardized source domain feature reference set. Compare the domain difference values ​​before and after adjustment. If the adjusted domain difference value does not decrease, reselect representative samples from the source domain feature reference set, or adjust the adjustment coefficients of the domain adaptation adjustment strategy, and perform adaptation adjustment on the transformed weak information region features again. If the adjusted domain difference value decreases, perform dimensional co-check on the adjusted features to ensure that the proportional relationship between the dimensions of the adjusted features matches the proportional relationship of the source domain training features.

[0125] After adjustment, domain offset verification is performed on the adjusted features. First, the adjusted features are standardized using the same standardization parameters as the source domain feature reference set (i.e., the mean and standard deviation of the source domain). Then, the domain difference value between the standardized adjusted features and the standardized source domain feature reference set is calculated using the same method as before. This new domain difference value is compared with the domain difference value before adjustment. If the new domain difference value has not decreased, it indicates that the adaptation adjustment has not effectively reduced the cross-domain difference. It is necessary to reselect representative samples of the source domain feature reference set (e.g., increase the number of samples or change the sampling method), or adjust the adjustment coefficients of the feature dimensions in the current domain adaptation adjustment strategy (e.g., increase the coefficients of important feature dimensions), and then perform adaptation adjustment on the transformed weak information region features again. If the new domain difference value has decreased, it indicates that the adjustment is effective. Then, a dimensional co-check is performed on the adjusted features. By fine-tuning the values ​​of each feature dimension, the proportional relationship between the dimensions in the adjusted features (e.g., the ratio of the maximum to the minimum value, the ratio of the means of different dimensions, etc.) is made as consistent as possible with the proportional relationship of the corresponding dimensions in the source domain training features, ensuring the consistency of the feature structure.

[0126] Step S1429: After completing the dimensionality co-check, output the domain-adapted features.

[0127] Once the adjusted features pass the domain offset verification and the proportional relationships of each dimension match those of the source domain training features, the cross-domain processing layer outputs the domain-adapted adjusted features. At this point, the features have been processed, and the difference between their statistical distribution and the source domain features is significantly reduced, making them more suitable for input into subsequent recognition modules.

[0128] Step S143: Input the domain-adapted features into the feature enhancement layer. The feature enhancement layer reads the feature dimension importance weight vector learned during the source domain training process, and performs element-wise multiplication of the weight vector with the domain-adapted features to generate the enhanced features.

[0129] The role of the feature enhancement layer is to highlight the feature dimensions that contribute more to the recognition task and suppress the interference of secondary feature dimensions. This layer reads the feature dimension importance weight vector learned during source domain training through feature importance evaluation (such as decision tree-based feature importance or gradient-based feature importance methods) from the model storage. The length of this feature dimension importance weight vector is the same as the number of feature dimensions, and each element corresponds to a feature dimension's importance weight value; the higher the importance of a dimension, the larger its corresponding weight value. This weight vector is then multiplied element-wise with the domain-adapted feature vector. That is, each element (feature value) in the feature vector is multiplied by the corresponding weight value in the weight vector, thereby amplifying the values ​​of important feature dimensions and relatively reducing the values ​​of secondary feature dimensions, generating enhanced features.

[0130] Step S144: Input the enhanced features into the type prediction layer, which is a fully connected network. It loads the network connection weights and bias parameters obtained from source domain training, performs linear combination and normalized exponential function calculation on the enhanced features, and outputs the recognition probability value corresponding to each mineral type.

[0131] The type prediction layer is the output layer of the model, typically consisting of one or more fully connected sublayers. This layer is loaded with the network connection weight matrix and bias vector parameters optimized through backpropagation during source domain training. The enhanced feature vector is input to the first fully connected sublayer, multiplied by its weight matrix, and then the bias vector is added, completing one linear combination. If multiple fully connected sublayers exist, the output of the previous sublayer becomes the input of the next, performing a new round of linear combination. The output vector of the last fully connected sublayer has a length equal to the number of mineral types to be identified. A normalized exponential function (Softmax function) is applied to this output vector, converting each element into a value between 0 and 1, with the sum of all elements being 1. These values ​​represent the probability of the input feature belonging to each mineral type.

[0132] Step S145: Sort the identification probability values, select the mineral type with the highest identification probability value as the prediction type of the sample to be identified in the weak information area, and record the identification probability value corresponding to the prediction type and the source domain parameter weights used by the model in the prediction process.

[0133] The identification probability values ​​of all mineral types output by the type prediction layer are sorted from largest to smallest. After sorting, the mineral type with the highest identification probability value is selected as the predicted type for the sample to be identified in the current weak information area. Simultaneously, the specific identification probability value corresponding to this predicted type, as well as the weight information of key source domain parameters used by the model throughout the prediction process, such as the linear transformation weights of the feature transfer layer, the adjustment coefficients of the cross-domain processing layer, and the importance weights of the feature enhancement layer, are recorded. This information is crucial for subsequent analysis of prediction results and evaluation of model transfer performance.

[0134] Step S146: Repeat the step of inputting the mineral completion features of the weak information area into the feature transfer layer of the transfer learning recognition model to sort the recognition probability values, and perform cross-domain recognition processing on each sample in the mineral sample set to be identified in the weak information area to generate a single sample recognition record for each sample.

[0135] For each sample in the set of mineral samples to be identified in the weak information area, the cross-domain identification processing flow from step S141 to step S145 is repeated. That is, the completed features of each sample are sequentially input into the feature transfer layer, cross-domain processing layer, feature enhancement layer, and type prediction layer to obtain its predicted type and recognition probability value, and the relevant parameters are recorded. An independent single sample identification record is generated for each sample, including information such as sample identifier, predicted type, recognition probability value, and source domain parameter weight identifier used.

[0136] Step S147: Integrate all single sample identification records to form a preliminary identification result of minerals in weak information areas, which includes the predicted type, identification probability value, and source domain parameter weight identifier for each sample to be identified.

[0137] The individual sample identification records of all samples to be identified are summarized and integrated according to the order of the samples in the set or the sample identifier to form a complete preliminary mineral identification result report for weak information areas. This preliminary mineral identification result report for weak information areas lists in detail the predicted mineral type, the corresponding identification probability value, and the identifier of the source domain parameter weights on which the prediction process is based (the specific parameter values ​​can be traced through the identifier).

[0138] Step S150: Select source domain samples from the source domain mineral sample set that are similar to the geological association elements of the weak information area, input the features of the source domain samples into the transfer learning recognition model to obtain the source domain sample recognition results, compare the source domain sample recognition results with the labeled types of the source domain samples to determine the transfer deviation, correct the preliminary mineral recognition results of the weak information area based on the transfer deviation, and generate the final mineral recognition results of the weak information area.

[0139] Step S151: Extract the overall features of geological correlation elements in the weak information area from the set of mineral samples to be identified in the weak information area, and form a geological feature template for the weak information area.

[0140] To find source domain samples with similar geological backgrounds to areas with weak information, it is first necessary to construct a geological feature template for these areas. From the set of mineral samples to be identified in the weak information areas, the geological correlation features (stratigraphic lithology, tectonic activity, and mineral assemblage) of all samples are extracted. Through statistical analysis of these features (such as calculating the mean, median, and mode of each feature dimension, and analyzing the distribution frequency of major feature types), a comprehensive summary of overall features reflecting the overall geological background and characteristics of the weak information areas is obtained. These features include major lithological assemblage types, dominant tectonic orientations, and common mineral assemblages. This overall summary of features forms the geological feature template for the weak information areas.

[0141] Step S152: Traverse each source domain sample in the source domain mineral sample set, extract the geological correlation element features of each source domain sample, calculate its similarity with the geological feature template of the weak information area, compare the similarity with the preset similarity threshold, and select the source domain samples whose similarity meets the preset similarity threshold requirements to form a source domain calibration sample group.

[0142] For each source region sample in the source region mineral sample set, extract its complete geological correlation element feature vector. Using a similar similarity calculation method as in step S127, calculate the similarity between the geological correlation element feature vector of the source region sample and the geological feature template of the weak information area. A preset similarity threshold is established, based on the diversity of source region samples and the typicality of geological features in the weak information area. Compare the calculated similarity value with the preset similarity threshold, and select source region samples with similarity values ​​greater than or equal to the threshold. These samples have a high similarity to the weak information area in terms of geological background, and are grouped into a source region calibration sample group for subsequent analysis of migration bias.

[0143] Step S153: Extract the source domain mineral features of each source domain sample in the source domain calibration sample group, input them into the transfer learning recognition model, perform recognition processing according to the cross-domain recognition process of weak information area samples, and generate the source domain sample recognition result of each source domain calibration sample.

[0144] Source domain mineral features are extracted one by one from the source domain calibration sample group. These features have the same dimension and structure as the mineral completion features in weak information areas. The extracted source domain mineral features are input into the transfer learning recognition model and processed strictly according to the cross-domain recognition process for weak information area samples. That is, the feature transfer layer performs linear transformation and nonlinear activation, the cross-domain processing layer calculates and adapts the domain differences, and the feature enhancement layer multiplies the importance weights. Finally, the type prediction layer outputs the recognition probability value and determines the predicted type. The above complete process is performed for each source domain calibration sample to generate a source domain sample recognition result containing the predicted type and the corresponding recognition probability value.

[0145] Step S154: Extract the labeled mineral type of each source domain sample in the source domain calibration sample group, compare it with the predicted type in the corresponding source domain sample identification result, and count the number of samples with prediction errors and the error types. The error types include type misjudgment and probability value deviation.

[0146] The labeled mineral type of each source domain sample is extracted from the metadata of the source domain calibration sample group. This labeled mineral type is a real type verified by actual exploration. The labeled mineral type is compared one by one with the predicted type in the source domain sample identification results obtained by the model. If the predicted type is completely inconsistent with the labeled type, it is judged as a type misclassification; if the predicted type is consistent with the labeled type, but the identification probability value is significantly lower than the preset confidence level (e.g., much lower than 0.8), it is judged as a probability value deviation. The number of samples with type misclassification and probability value deviation, as well as the total number of incorrect prediction samples, are counted separately.

[0147] Step S155: For each incorrectly predicted sample, calculate the identification deviation value. The identification deviation value is the difference between the standard probability value corresponding to the labeled type and the model output identification probability value. Summarize the migration deviation pattern based on the identification deviation values ​​of all incorrect samples.

[0148] Step S1551: Extract the labeled mineral type of each erroneous sample from the predicted erroneous samples of the source domain calibration sample group, and set a standard probability value for each labeled type. The standard probability value is set as the baseline probability that the labeled type is the correct result.

[0149] For all samples in the source domain calibration sample set that are judged to be incorrect predictions, their labeled mineral types are extracted. According to the general requirements of mineral identification tasks, a standard probability value is set for each labeled type. This standard probability value represents the ideal recognition probability value that the model should output when the sample is indeed of that type, and serves as a benchmark probability for measuring the deviation of the model output.

[0150] Step S1552: Extract the model output recognition result for each erroneous sample and obtain the recognition probability value of the labeled type in the model output.

[0151] Even if the source domain calibration sample is incorrectly predicted by the model, the probability vector output by the model still contains the probability value of the sample's true type (labeled type). The recognition probability value corresponding to the labeled type is extracted from the model's output recognition result for each incorrect sample.

[0152] Step S1553: Calculate the recognition deviation value for each erroneous sample. The recognition deviation value is equal to the standard probability value minus the recognition probability value output by the model. A positive recognition deviation value indicates that the model's recognition probability of the correct type is too low, and a negative recognition deviation value indicates that the model's recognition probability of the correct type is too high.

[0153] For each incorrectly predicted sample, the labeled type recognition probability value output by the model is subtracted from the set standard probability value. The result is the recognition bias value. If the recognition bias value is positive, it means that the model's recognition probability of the correct type is lower than the ideal level; if it is negative, it means that the model's recognition probability of the correct type is higher than the ideal level (the above situation is rare in type misjudgment, but may occur in probability value bias).

[0154] Step S1554: Classify the erroneous samples according to the labeled mineral type, group the erroneous samples of the same labeled type together to form multiple groups of type deviation sample groups.

[0155] All predicted error samples are grouped according to their labeled mineral types. For example, all error samples labeled as "galena ore bodies" are grouped into one group, and all error samples labeled as "sphalerite ore bodies" are grouped into another group, and so on, forming multiple type bias sample groups, each corresponding to a specific mineral type.

[0156] Step S1555: For each group of type deviation samples, calculate the average and variance of the identification deviation values ​​of all erroneous samples in the group, and determine the central range and dispersion of the identification deviation of that type.

[0157] Statistical analysis was performed on all identification deviation values ​​within each type of deviation sample group to calculate their mean and variance. The mean reflects the overall direction and average magnitude of the identification deviation for that mineral type, i.e., the range of concentration; the variance reflects the degree of dispersion of the identification deviation values ​​around the mean, with a larger variance indicating greater instability in the magnitude of the deviation.

[0158] Step S1556: Analyze the geological correlation features of samples in each group of type deviation samples, find the common geological feature attributes of the deviation sample group, and determine whether there are geological attributes that cause identification bias.

[0159] A thorough analysis of the geological features of all samples in each type of biased sample group was conducted, comparing the differences in geological characteristics between these samples and the correctly predicted source domain samples of the same type. Special attention was paid to whether there were any shared geological features, such as specific rock assemblages, unique tectonic locations, or anomalous mineral assemblages, as these shared attributes could be the cause of systematic identification bias in the model for that type.

[0160] Step S1557: Based on the deviation statistics of different types of deviation sample groups and the associated geological feature attributes, summarize the migration deviation rules, which include the correspondence between the deviation direction and magnitude of mineral types and geological feature attributes.

[0161] By synthesizing the statistical results (mean and variance) of the deviations from various types of deviation sample groups and analyzing the associated geological characteristics, a migration deviation pattern is summarized. This migration deviation pattern clearly describes, under what geological characteristic conditions, the direction (too high or too low) and approximate degree (range of deviation value) of the model's identification deviation of mineral types, thus establishing the correspondence between mineral type, deviation direction, deviation magnitude, and geological characteristic attributes.

[0162] Step S1558: Verify the summarized migration deviation rules. Select erroneous samples from the source domain calibration sample group that were not included in the rule summary, and analyze whether the identification deviation of the sample conforms to the summarized migration deviation rules. If it conforms to the summarized migration deviation rules, the migration deviation rules are determined to be valid. If there are samples that do not conform to the summarized migration deviation rules, re-analyze the geological characteristics of the sample and supplement or adjust the content of the migration deviation rules.

[0163] A portion of the incorrectly predicted samples from the source domain calibration sample group, which were not included in the summary of migration deviation patterns, are reserved as validation samples. The geological characteristics of these validation samples are compared with the summarized migration deviation patterns to check whether the direction and magnitude of the identified deviations conform to the predicted patterns. If all validation samples conform to the patterns, the migration deviation patterns are considered valid; if any samples do not conform, their geological characteristics need to be re-analyzed to identify potentially omitted or incorrectly associated factors in the patterns, thereby supplementing or adjusting the content of the migration deviation patterns to ensure their accuracy and universality.

[0164] Step S1559: Form the final migration deviation pattern document, which includes deviation parameters, associated geological feature attributes, and pattern verification results for each mineral type.

[0165] The verified and adjusted migration deviation patterns were compiled into a formal migration deviation pattern document. The document details the average deviation value, deviation variance, and other deviation parameters for each mineral type, the specific associated geological characteristics that led to the deviation, and the verification results, such as the conformity rate of the verification samples obtained during the pattern verification process.

[0166] Step S156: Construct a deviation correction model based on the migration deviation law. The deviation correction model includes correction coefficients for different error types. The values ​​of the correction coefficients are determined by referencing the similarity between the geological correlation elements of the source domain calibration sample and the weak information area.

[0167] A deviation correction model is constructed based on the migration deviation pattern document. The model sets correction coefficients for two error types: type misjudgment and probability value deviation. For probability value deviation, the correction coefficient is determined based on the direction and magnitude of the deviation of the mineral type under specific geological feature attributes; for type misjudgment, the correction coefficient is used to adjust the recognition probability weights between different mineral types. When determining the specific values ​​of the correction coefficients, the similarity between the geological association elements of the source domain calibration sample and the geological feature template of the weak information area is fully considered. The higher the similarity, the closer the value of the correction coefficient is to the deviation parameters of the source domain calibration sample.

[0168] Step S157: Input the preliminary mineral identification results of the weak information area into the deviation correction model. The model calls the matching correction coefficient to adjust the identification probability value according to the predicted type and corresponding identification probability value of each sample to be identified. If a new highest probability type appears after adjustment, the predicted type is updated.

[0169] The preliminary mineral identification results for weak information areas are input into the bias correction model. For each sample to be identified, the model first extracts its predicted type, corresponding identification probability value, and its own geological correlation features. Based on the geological correlation features of the sample, the most similar geological feature attributes are matched in the migration bias pattern document, and then the corresponding correction coefficient is retrieved from the bias correction model. The correction coefficients are used to adjust the identification probability values ​​of all mineral types for the sample. For example, for mineral types with positive bias under the current geological features, their identification probability values ​​are increased; for types with negative bias, their identification probability values ​​are decreased. After adjustment, the identification probability values ​​of all mineral types are re-sorted. If a new highest probability type appears, the predicted type of the sample is updated to this new type.

[0170] Step S158: Perform calibration verification on the corrected identification results, select the source domain sample in the source domain calibration sample group that is most similar to the geological features of the sample to be identified in the weak information area, and analyze whether the corrected identification results of the sample to be identified are consistent with the identification logic of the source domain sample.

[0171] To ensure the reliability of the corrected identification results, calibration verification is performed on the corrected results for each sample to be identified in a weak information region. Specifically, the source domain sample most similar to the geological features of the sample to be identified is selected from the source domain calibration sample group (determined by calculating feature similarity). The corrected predicted type and identification probability distribution of the sample to be identified are analyzed and compared with the identification results of the most similar source domain sample in the model (especially the consistency between the labeled type and the predicted type, and the reasonableness of the identification probability values). This checks whether the identification logic of the two is consistent, such as whether similar identification conclusions are obtained due to similar geological features.

[0172] Step S159: If the calibration verification meets the requirements, retain the corrected identification results; if the calibration verification does not meet the requirements, adjust the correction coefficient of the deviation correction model and re-correct the preliminary identification results of minerals in the weak information area.

[0173] If the corrected identification result of the sample to be identified is consistent with the identification logic of the most similar source domain sample, and the identification probability value is within a reasonable range, then the calibration verification is deemed to meet the requirements, and the corrected identification result is retained. If the calibration verification does not meet the requirements, it indicates that the correction coefficients of the current deviation correction model may not be accurate enough, and it is necessary to return to step S156. Based on the problems found in this calibration verification (such as over-correction or under-correction), the correction coefficients of the corresponding geological feature attributes and mineral types in the deviation correction model are adjusted. Then, the adjusted model is used to re-correct the preliminary identification results of minerals in the weak information area, and the calibration verification is performed again until the correction results of all samples pass the verification.

[0174] Step S1510: Integrate all the correction results that meet the calibration and verification requirements to generate the final identification result of minerals in the weak information area. The final identification result includes the corrected prediction type, adjusted probability value, and deviation correction basis for each sample to be identified.

[0175] The corrected identification results of all weak information area samples that have passed calibration and verification are summarized and integrated. The final mineral identification result of the weak information area is a structured report, which details the corrected predicted mineral type, the adjusted identification probability value, and the specific migration deviation rules and correction coefficient values ​​on which the deviation correction was based for each sample, ensuring the traceability and reliability of the results.

[0176] Furthermore, Figure 2 A schematic diagram of the hardware structure of a transfer learning-based intelligent mineral identification system 100 for implementing the method provided in the embodiments of this application is shown. Figure 2 As shown, the intelligent mineral identification system 100 based on transfer learning in weak information areas may include at least one processor 102 (the processor 102 may be, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, a transmission device 106 for communication functions, and a controller 108. Those skilled in the art will understand that... Figure 2 The structure shown is for illustrative purposes only and does not limit the structure of the intelligent mineral identification system 100 for weak information areas based on transfer learning. For example, the intelligent mineral identification system 100 for weak information areas based on transfer learning may also include... Figure 2 The more or fewer components shown, or having the same Figure 2 The different configurations shown.

[0177] The memory 104 can be used to store software programs and modules for application software, such as the program instructions corresponding to the method embodiments described above in this application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the above-described intelligent identification method for mineral resources in weak information areas based on transfer learning. The transmission device 106 is used to acquire or send data via a network.

[0178] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

Claims

1. A method for intelligent identification of mineral resources in weak information areas based on transfer learning, characterized in that, The method includes: Obtain a source domain mineral sample set and a weak information area mineral sample set to be identified. The source domain mineral sample set contains source domain mineral features with labeled mineral types, and the weak information area mineral sample set to be identified contains original features of weak information area minerals without labeled mineral types. Geological correlation elements are extracted from the source domain mineral features and the original mineral features of the weak information area. Based on the correspondence between the source domain geological correlation elements and the weak information area geological correlation elements, a cross-domain feature mapping is constructed. The original mineral features of the weak information area are then transferred and adapted to generate intermediate mineral features of the weak information area that are adapted to the source domain feature dimensions. Identify the sparse feature dimensions in the intermediate features of minerals in the weak information region, call the feature distribution pattern of the corresponding dimension in the source domain mineral features, and complete the sparse feature dimensions through the feature transfer mechanism of transfer learning to generate complete features of minerals in the weak information region. The mineral completion features of the weak information area are input into a pre-trained transfer learning recognition model. The transfer learning recognition model loads the feature processing parameters obtained from source domain training and performs cross-domain recognition processing on the mineral completion features of the weak information area to generate preliminary mineral recognition results for the weak information area. Select source domain samples from the source domain mineral sample set that are similar to the geological association elements of the weak information area, input the features of the source domain samples into the transfer learning recognition model to obtain the source domain sample recognition results, compare the source domain sample recognition results with the labeled types of the source domain samples to determine the transfer deviation, correct the preliminary mineral recognition results of the weak information area based on the transfer deviation, and generate the final mineral recognition results of the weak information area. The process involves extracting geologically related elements from the source domain mineral features and from the original mineral features of the weak-information area, constructing a cross-domain feature mapping based on the correspondence between the source domain geologically related elements and the weak-information area geologically related elements, performing transfer adaptation processing on the original mineral features of the weak-information area, and generating intermediate mineral features of the weak-information area that are adapted to the feature dimensions of the source domain, including: Geological correlation elements are extracted from the source region mineral characteristics. The geological correlation elements include the stratigraphic and lithological characteristics, tectonic activity characteristics, and mineral symbiotic assemblage characteristics of the source region mineral area. Each geological correlation element carries a feature identifier corresponding to the source region mineral type. Geological correlation elements are extracted from the original mineral characteristics of the weak information area. The geological correlation elements include the stratigraphic lithology, tectonic activity, and mineral symbiotic assemblage characteristics of the area where the minerals are located. Each geological correlation element is associated with the collection scene information of the weak information area sample. Cross-domain correlation analysis was conducted on the stratigraphic lithological characteristics of the source domain and the weak information area. The differences in the characteristic performance of the same lithological type in the source domain and the weak information area were calculated, and the cross-domain mapping weights of the lithological characteristics were assigned based on the difference values. Using the same method, the cross-domain performance differences of tectonic activity characteristics and mineral symbiotic assemblage characteristics between the source domain and the weak information region were calculated, and cross-domain mapping weights were assigned to the corresponding characteristics. By integrating the cross-domain mapping weights of stratigraphic lithology characteristics, tectonic activity characteristics, and mineral symbiotic assemblage characteristics, a cross-domain feature mapping matrix between the source domain and the weak information area is constructed. Each element in the cross-domain feature mapping matrix corresponds to a set of mapping relationships and weights of geological correlation elements between the source domain and the weak information area. The original mineral features of the weak information area are input into the cross-domain feature mapping matrix. The feature values ​​of each geological associated element in the weak information area are migrated and adjusted according to the corresponding mapping weights so that the adjusted feature values ​​conform to the representation rules of the source domain features. The migration fit verification is performed on the adjusted weak information area features. The similarity between the adjusted weak information area geological associated element features and the source domain similar geological associated element features is calculated, and the similarity is compared with the preset migration fit threshold. If the similarity is lower than the preset migration adaptation threshold, the cross-domain mapping weight of the corresponding geological associated elements will be readjusted, and the feature values ​​of the geological associated elements in the weak information area will be migrated and adjusted again. If the similarity reaches the preset migration adaptation threshold, cross-domain collaborative integration is carried out on all the adjusted geological associated element features. By adjusting the numerical distribution of the geological associated element features in the weak information area, the feature association relationship between different geological associated elements conforms to the association relationship of the source domain features. Based on the collaboratively integrated features, intermediate features of mineral resources in weak information areas are generated that are adapted to the feature dimensions of the source domain. Each geological correlation element of the intermediate features of mineral resources in weak information areas carries a weight identifier for source domain migration adaptation.

2. The intelligent identification method for mineral resources in weak information areas based on transfer learning according to claim 1, characterized in that, The process involves identifying sparse feature dimensions in the intermediate features of minerals in the weak information region, invoking the feature distribution patterns of corresponding dimensions in the source domain mineral features, and completing the sparse feature dimensions through a feature transfer mechanism of transfer learning to generate completed features for minerals in the weak information region. This includes: The number of effective features in each feature dimension of the intermediate features of minerals in the weak information area is counted, the proportion of effective features is calculated, and feature dimensions with an effective feature proportion lower than a preset sparsity threshold are marked as sparse feature dimensions, thus forming a list of sparse feature dimensions. Extract source domain feature data corresponding to each dimension in the sparse feature dimension list from the source domain mineral features, and analyze the numerical distribution pattern of the source domain feature data. The numerical distribution pattern includes the concentration range of feature values, distribution density, and correlation between adjacent dimensions. For each sparse feature dimension, a transfer completion model is constructed based on the feature distribution pattern of the corresponding dimension in the source domain. The transfer completion model loads the feature distribution parameters of the source domain and has the ability to transfer the distribution pattern of the source domain to the sparse dimension of the weak information region. The existing effective feature data in the sparse feature dimension of the weak information region is input into the transfer completion model. The transfer completion model adapts the effective feature data to the distribution of the source domain based on the distribution law of the source domain, and determines the feature distribution trend of the sparse dimension of the weak information region. Based on the aforementioned feature distribution trend, numerical prediction is performed on the location of missing data in the sparse feature dimension of the weak information region. During the prediction process, the correlation between adjacent dimensions in the source domain features is referenced to ensure that the predicted values ​​are consistent with the existing effective feature data in the weak information region. The predicted values ​​are filled into the missing data positions to complete the initial completion of the sparse feature dimensions, thus obtaining the initial completed features. Perform a transfer consistency check on the initially completed features, calculate the matching degree between the completed weak information region features and the corresponding dimension features of the source domain in terms of distribution patterns, and compare the matching degree with a preset consistency threshold. If the matching degree is lower than the preset consistency threshold, the source domain parameter weights of the transfer completion model are adjusted, and the effective feature data of the sparse feature dimension of the weak information region is re-inputted into the model for numerical prediction. If the matching degree reaches the preset consistency threshold, cross-dimensional collaborative verification is performed on the completion results of all sparse feature dimensions to analyze the correlation between different sparse dimensions after completion, so that the correlation conforms to the correlation rules of the source domain features. After collaborative verification, the sparse feature dimensions of the completed feature are integrated with the original non-sparse feature dimensions to generate mineral completion features for weak information regions. Each dimension of the mineral completion features for weak information regions carries a migration identifier of the source domain distribution pattern.

3. The intelligent identification method for mineral resources in weak information areas based on transfer learning according to claim 1, characterized in that, The pre-trained transfer learning recognition model includes a feature transfer layer, a cross-domain processing layer, and a type prediction layer. The feature transfer layer loads the feature processing parameters obtained from training in the source domain. The cross-domain processing layer is used to eliminate the feature domain differences between the source domain and the weak information region. The type prediction layer is used to output the mineral type recognition result. The step involves inputting the mineral completion features of the weak information area into a pre-trained transfer learning recognition model. The transfer learning recognition model loads feature processing parameters obtained from source domain training and performs cross-domain recognition processing on the mineral completion features of the weak information area to generate preliminary mineral identification results for the weak information area, including: The mineral completion features of the weak information area are input into the feature transfer layer of the transfer learning recognition model. The feature transfer layer calls the linear transformation parameters and activation function parameters obtained from the source domain training to perform linear transformation and nonlinear activation operations on the weak information area completion features to generate the transformed features. The transformed features are input into the cross-domain processing layer, which calculates the difference measure between the transformed features and the source domain training feature set extracted from the model storage in terms of statistical distribution. Based on the difference measure, the transformed features are numerically adjusted through a trainable adaptation network to generate domain-adapted features. The domain-adapted features are input into the feature enhancement layer. The feature enhancement layer reads the feature dimension importance weight vector learned during the source domain training process, and performs element-wise multiplication of the weight vector with the domain-adapted features to generate the enhanced features. The enhanced features are input into the type prediction layer, which is a fully connected network. It loads the network connection weights and bias parameters obtained from source domain training, performs linear combination and normalized exponential function calculation on the enhanced features, and outputs the recognition probability value corresponding to each mineral type. The identification probability values ​​are sorted, and the mineral type with the highest identification probability value is selected as the prediction type of the sample to be identified in the weak information area. The identification probability value corresponding to the prediction type and the source domain parameter weights used by the model in the prediction process are recorded. Repeat the steps of inputting the mineral completion features of the weak information area into the feature transfer layer of the transfer learning recognition model to sort the recognition probability values, and perform cross-domain recognition processing on each sample in the mineral sample set to be identified in the weak information area to generate a single sample recognition record for each sample. By integrating all individual sample identification records, a preliminary mineral identification result for weak information areas is formed, which includes the predicted type, identification probability value, and source domain parameter weight identifier for each sample to be identified.

4. The intelligent identification method for mineral resources in weak information areas based on transfer learning according to claim 1, characterized in that, The process involves selecting source domain samples from the source domain mineral sample set that are similar to geologically related elements in the weak information area, inputting the features of the source domain samples into the transfer learning recognition model to obtain source domain sample recognition results, comparing the source domain sample recognition results with the labeled types of the source domain samples to determine the transfer bias, correcting the preliminary mineral identification results in the weak information area based on the transfer bias, and generating the final mineral identification results in the weak information area, including: Extract the overall features of geologically related elements in the weak information area from the set of mineral samples to be identified in the weak information area, and form a geological feature template for the weak information area; Traverse each source domain sample in the source domain mineral sample set, extract the geological correlation element features of each source domain sample, calculate its similarity with the geological feature template of the weak information area, compare the similarity with the preset similarity threshold, and select the source domain samples whose similarity meets the preset similarity threshold requirements to form a source domain calibration sample group. Extract the source domain mineral features of each source domain sample in the source domain calibration sample group, input them into the transfer learning recognition model, perform recognition processing according to the cross-domain recognition process of weak information area samples, and generate the source domain sample recognition result of each source domain calibration sample; Extract the labeled mineral type of each source domain sample in the source domain calibration sample group, compare it with the predicted type in the corresponding source domain sample identification result, and count the number of samples with prediction errors and the error types. The error types include type misjudgment and probability value deviation. For each incorrectly predicted sample, the identification deviation value is calculated. The identification deviation value is the difference between the standard probability value corresponding to the labeled type and the identification probability value output by the model. Based on the identification deviation values ​​of all incorrect samples, the migration deviation pattern is summarized. Based on the aforementioned migration deviation pattern, a deviation correction model is constructed. The deviation correction model includes correction coefficients for different error types, and the values ​​of the correction coefficients are determined with reference to the similarity between the geological correlation elements of the source domain calibration sample and the weak information area. The preliminary mineral identification results of the weak information area are input into the deviation correction model. The model calls the matching correction coefficient to adjust the identification probability value according to the predicted type and corresponding identification probability value of each sample to be identified. If a new highest probability type appears after adjustment, the predicted type is updated. The corrected identification results are calibrated and verified. The source domain sample that is most similar to the geological features of the sample to be identified in the weak information area is selected from the source domain calibration sample group. The corrected identification results of the sample to be identified are analyzed to see if they are consistent with the identification logic of the source domain sample. If the calibration verification meets the requirements, the corrected identification results are retained; if the calibration verification does not meet the requirements, the correction coefficient of the deviation correction model is adjusted, and the preliminary identification results of minerals in the weak information area are corrected again. By integrating all the corrected results that meet the calibration and verification requirements, the final identification result of minerals in the weak information area is generated. The final identification result includes the corrected prediction type, adjusted probability value, and deviation correction basis for each sample to be identified.

5. The intelligent identification method for mineral resources in weak information areas based on transfer learning according to claim 1, characterized in that, The cross-domain correlation analysis of stratigraphic lithological characteristics between the source domain and the weak information area calculates the differences in characteristic performance of the same lithological type in the source domain and the weak information area, and assigns cross-domain mapping weights to lithological characteristics based on the difference values, including: All lithological types are extracted from the mineral characteristics of the source region to establish a source region lithological type library. Each lithological type corresponds to a set of characteristic performance data, which includes the proportion of lithological components and structural density characteristics. All lithological types are extracted from the original mineral characteristics of weak information areas to establish a lithological type library for weak information areas. Each lithological type also corresponds to a set of characteristic performance data, which includes the proportion of lithological components and structural density characteristics. Type matching is performed between the source domain lithology type database and the weak information area lithology type database to identify lithology types with the same name or similar attributes, forming cross-domain lithology matching pairs. For each cross-domain lithology matching pair, multiple sets of characteristic performance data of that lithology type in the source domain are selected, and the mean and standard deviation of the characteristic data are calculated to determine the standard performance range of the lithology characteristics in the source domain. Using the same method, the mean and standard deviation of the characteristic performance data of that lithology type in the weak information area are calculated to determine the actual performance range of the lithology characteristics in the weak information area. Calculate the difference between the standard performance range and the actual performance range, where the difference is the ratio of the absolute difference between their means to the sum of their standard deviations. Set thresholds for different values, including a first threshold and a second threshold, where the first threshold is less than the second threshold. If the difference value is less than the first threshold, a first mapping weight is assigned to the cross-domain lithology matching pair; if the difference value is greater than or equal to the first threshold and less than the second threshold, a second mapping weight is assigned to the cross-domain lithology matching pair; if the difference value is greater than or equal to the second threshold, a third mapping weight is assigned to the cross-domain lithology matching pair; the first mapping weight is greater than the second mapping weight, and the second mapping weight is greater than the third mapping weight. Select lithological types from the source domain lithological type library that are similar to but not completely matched with the lithological type attributes of the weak information area, calculate their characteristic difference values ​​with the lithological types of the weak information area, and assign auxiliary mapping weights according to the correspondence between the above difference value classification threshold and mapping weights. The value of the auxiliary mapping weight is lower than the first mapping weight, second mapping weight or third mapping weight of the corresponding difference value interval. The initial cross-domain mapping weights and auxiliary mapping weights are weighted and fused. During the fusion process, the lithological type data of the weak information area is taken into account. If the similarity between the data collection scene and the source domain meets the preset similarity requirements, the proportion of the initial mapping weights in the fusion is increased. Based on the fused weight values, the cross-domain mapping weight of lithological characteristics for each cross-domain lithology matching pair is determined. The sum of the weight values ​​is set to a fixed total weight value to ensure a balanced weight distribution among different lithology types. Record the feature difference values ​​and weight allocation process for each cross-domain lithology matching pair to form a cross-domain mapping weight table of lithology features.

6. The intelligent identification method for mineral resources in weak information areas based on transfer learning according to claim 2, characterized in that, The process of performing a transfer consistency check on the initially completed features involves calculating the matching degree between the completed weak information region features and the corresponding dimension features of the source domain in terms of distribution patterns. This matching degree is then compared to a preset consistency threshold. If the matching degree is lower than the preset consistency threshold, the source domain parameter weights of the transfer completion model are adjusted, and the effective feature data of the sparse feature dimension of the weak information region is re-inputted into the model for numerical prediction. This includes: Source domain feature data corresponding to sparse feature dimensions are extracted from source domain mineral features. Statistical methods are used to analyze the distribution pattern of source domain feature data to generate source domain feature distribution curves. The source domain feature distribution curves include parameters such as numerical frequency, peak position, and distribution width. Extract the corresponding sparse dimension completion feature data from the initial completion features, and use the same statistical method to generate completion feature distribution curves so that the statistical dimensions of the two distribution curves are consistent. The similarity between the source domain feature distribution curve and the completed feature distribution curve is calculated. The similarity is obtained by comprehensively calculating the percentage of overlapping area of ​​the curves, the peak position deviation, and the distribution width deviation. The result is the matching degree. The matching degree is compared with a preset consistency threshold. If the matching degree meets the preset consistency threshold requirement, the preliminary completed feature is determined to have passed the migration consistency check. If the matching degree does not meet the preset consistency threshold requirement, the difference position of the two distribution curves is analyzed, and the average difference between the completed feature value and the source domain feature value in each numerical interval of the completed feature distribution curve is calculated. The interval with the average difference higher than the preset difference threshold is determined as the numerical interval that needs to be adjusted. For numerical ranges that require key adjustments, adjust the weights of the corresponding source domain parameters in the transfer completion model, reduce the influence weights of the source domain parameters within the numerical range, and increase the weights of the source domain parameters that match the existing effective feature data in the weak information region. The adjustment range is referenced to the ratio of the average difference of the numerical range to the preset difference threshold; the higher the ratio, the greater the adjustment range. After adjusting the parameter weights, the effective feature data of the sparse feature dimension of the weak information region is re-input into the transfer completion model. The transfer completion model re-predicts the missing data position based on the adjusted parameter weights, generates a distribution curve for the re-predicted completed feature data, and calculates the matching degree between the distribution curve and the source domain feature distribution curve. Repeat the steps of adjusting the weights of the corresponding source domain parameters in the migration completion model, re-inputting the effective feature data into the migration completion model for numerical prediction, and calculating the matching degree between the new distribution curve and the source domain feature distribution curve, until the matching degree of the completed features meets the preset consistency threshold requirement.

7. The intelligent identification method for mineral resources in weak information areas based on transfer learning according to claim 3, characterized in that, The transformed features are input into a cross-domain processing layer. This layer calculates a statistical difference metric between the transformed features and the source domain training feature set extracted from the model storage. Based on this difference metric, a trainable adaptation network is used to numerically adjust the transformed features, generating domain-adapted features. This process includes: The weak information region features are obtained from the feature transfer layer of the transfer learning recognition model. At the same time, the source domain training feature sample set stored in the model is extracted, and representative samples from the source domain training feature sample set are selected to form the source domain feature reference set. The source domain feature reference set and the transformed weak information region features are standardized to eliminate the dimensional differences between different feature dimensions. Calculate the feature distance between the standardized weak information region features and the feature reference set of each sample in the standardized source domain feature reference set, and obtain a set of feature distance values ​​using the distance calculation method; The domain difference value is calculated based on the feature distance value, and the domain difference value is the average of all feature distance values. Set a grading standard for domain difference values, including a first-level threshold, a second-level threshold, and a third-level threshold. The first-level threshold is lower than the second-level threshold, and the second-level threshold is lower than the third-level threshold. If the domain difference value is less than the first-level threshold of the domain difference value, then the first domain adaptation adjustment strategy is applied, and the adjustment range of the first domain adaptation adjustment strategy is the minimum adjustment range; if the domain difference value is greater than or equal to the first-level threshold of the domain difference value and less than the second-level threshold of the domain difference value, then the second domain adaptation adjustment strategy is applied, and the adjustment range of the second domain adaptation adjustment strategy is the medium adjustment range; if the domain difference value is greater than or equal to the second-level threshold of the domain difference value and less than the third-level threshold of the domain difference value, then the third domain adaptation adjustment strategy is applied, and the adjustment range of the third domain adaptation adjustment strategy is the maximum adjustment range. Based on the level of the domain difference value of the weak information region features after transformation, the corresponding domain adaptation adjustment strategy is invoked. The domain adaptation adjustment strategy includes the adjustment coefficient of the feature dimension, and the adjustment coefficient is determined based on the dimensional importance of the source domain training features. The values ​​of each dimension of the transformed weak information region features are adapted and adjusted according to the adjustment coefficient, while maintaining the correlation between the feature dimensions during the adjustment process; The adjusted features are validated by domain offset and then standardized using the same standardization parameters as the source domain feature reference set. The domain difference between the standardized adjusted features and the standardized source domain feature reference set is then calculated. The domain difference before and after adjustment is compared. If the adjusted domain difference does not decrease, representative samples from the source domain feature reference set are reselected, or the adjustment coefficients of the domain adaptation strategy are adjusted, and the transformed weak information region features are adapted again. If the adjusted domain difference decreases, dimensional co-checking is performed on the adjusted features to ensure that the proportional relationship between the dimensions of the adjusted features matches the proportional relationship of the source domain training features. After completing the dimensional collaboration check, the output domain is adjusted to reflect the features.

8. The intelligent identification method for mineral resources in weak information areas based on transfer learning according to claim 4, characterized in that, For each incorrectly predicted sample, a recognition bias value is calculated. This recognition bias value is the difference between the standard probability value corresponding to the labeled type and the model's output recognition probability value. Based on the recognition bias values ​​of all incorrect samples, the transfer bias pattern is summarized, including: Extract the labeled mineral type of each erroneous sample from the predicted erroneous samples of the source domain calibration sample group, and set a standard probability value for each labeled type. The standard probability value is set as the baseline probability that the labeled type is the correct result. Extract the model output recognition result for each erroneous sample and obtain the recognition probability value of the labeled type in the model output; Calculate the identification bias value for each incorrect sample. The identification bias value is equal to the standard probability value minus the identification probability value output by the model. A positive identification bias value indicates that the model has a low probability of identifying the correct type, and a negative identification bias value indicates that the model has a high probability of identifying the correct type. The erroneous samples are classified according to the labeled mineral type, and the erroneous samples of the same labeled type are grouped together to form multiple groups of type deviation sample groups; For each group of type-biased samples, calculate the average and variance of the identification deviation values ​​of all erroneous samples within the group to determine the central tendency and dispersion of the identification deviation for that type. Analyze the geological features of samples in each group of type deviation samples, identify the common geological features of the deviation sample group, and determine whether there are geological attributes that cause identification bias. Based on the deviation statistics of different types of deviation sample groups and the associated geological characteristic attributes, the migration deviation pattern is summarized. The migration deviation pattern includes the correspondence between the deviation direction and magnitude of mineral types and geological characteristic attributes. To verify the summarized migration deviation patterns, select erroneous samples from the source domain calibration sample group that were not included in the pattern summary, and analyze whether the identification deviation of these samples conforms to the summarized migration deviation patterns. If they conform to the summarized migration deviation patterns, the migration deviation patterns are determined to be valid. If there are samples that do not conform to the summarized migration deviation patterns, reanalyze the geological characteristics of these samples and supplement or adjust the content of the migration deviation patterns. The final migration deviation pattern document is generated, which includes deviation parameters, associated geological feature attributes, and pattern verification results for each mineral type.

9. A mineral intelligent identification system for weak information areas based on transfer learning, characterized in that, The device includes a processor and a readable storage medium storing a program that, when executed by the processor, implements the intelligent identification method for mineral resources in weak information areas based on transfer learning as described in any one of claims 1-8.

Images