A Physically Constrained AI-Interpreted Method for SQUID Magnetic Gradient Target Recognition and Target Area Recommendation in Mining Areas
By constructing a class probability model trained with pixel-level physical feature stacks and pseudo-label samples, the problem of SQUID magnetic gradient inversion results being difficult to convert into target area recommendations was solved, achieving stable target area recommendations and target recognition, and improving the interpretability and verifiability of the results.
Patent Information
- Application Number
- CN202611018989.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-09
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to reliably convert SQUID magnetic gradient inversion results into verifiable target categories and target area recommendations for mining areas, especially when training without using ground truth labels for the current survey area. This makes it difficult to effectively utilize category probability, entropy ambiguity, and in-distribution reliability for target area recommendations.
By constructing a pixel-level physical feature stack, pseudo-label samples that do not use the ground truth labels of the current survey area are generated. A category probability model is trained, and the category probability, entropy ambiguity, and distribution reliability of underground objects are calculated to generate target area recommendation results.
It improves the interpretability and verifiability of target recognition results, reduces the risk of confusion between training and validation boundaries, and can output target area recommendation results containing category, coordinates, priority score and recommendation reason, which facilitates subsequent manual review and validation deployment.
Smart Images

Figure CN123113012A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mineral exploration and geophysical data processing technology, and in particular to a method for identifying mining targets and recommending target areas based on SQUID magnetic gradient physical inversion results. Background Technology
[0002] The SQUID magnetic gradient system can acquire information on weak magnetic fields or magnetic gradients in mining areas. After magnetostatic inversion, it can generate results such as underground magnetic sources, source depths, source densities, or physical fitting maps. Regarding the interpretation of inversion results and the recommendation of target areas in mining areas, existing technical approaches mainly include the following categories: 1. Technical personnel examine magnetic field maps, gradient maps, inversion source distributions, or 3D anomalies, and then perform manual interpretation in conjunction with geological, geochemical, remote sensing, and engineering data. This approach preserves expert experience, but the interpretation criteria are affected by personnel experience and data completeness, making it difficult to consistently output category probabilities, reliability, ambiguity, and recommendation rationales for the same anomaly area.
[0003] 2. Convert geophysical inversion results into subsurface property models or anomaly displays, and then determine the likelihood of mineralization based on electromagnetic, gravity, magnetic, or multiphysics field properties. This approach focuses on restoring subsurface properties or the spatial location of anomalies, and typically does not organize the SQUID magnetic gradient tensor invariants, equivalent magnetic sources, source densities, shallow source densities, residual structures, and posterior source densities into a pixel-level physical feature stack that can be probabilistically interpreted.
[0004] 3. Inputting magnetic gradient maps or anomaly maps into machine learning models for image classification or target recognition. If this approach lacks physical inversion products and geological prior constraints, it is prone to learning noise textures, missing morphologies, or synthetic data biases; if the ground truth labels of the current survey area are used directly for training, the boundary between training and validation will be unclear, making it difficult to demonstrate the generalization ability of new survey areas.
[0005] 4. Existing methods for target area recommendation mostly rely on manual delineation or screening using a single probability threshold, making it difficult to simultaneously utilize the probability of valuable targets, the probability of non-value or interference targets, class ambiguity, and distribution reliability. Furthermore, the area of connected regions and local peak splitting rules are not fully incorporated into the ranking process, thus hindering the formation of verifiable reasons for target area recommendations.
[0006] Therefore, a processing method is needed that can take the SQUID magnetic gradient physical inversion results, generate underground object category probabilities without training using the current survey area ground truth labels, and further transform the category probabilities, entropy ambiguity, and in-distribution reliability into target area recommendation results. Summary of the Invention
[0007] The purpose of this invention is to address the problem that SQUID magnetic gradient inversion products are difficult to reliably transform into verifiable target categories and target area recommendation results. This invention provides a SQUID magnetic gradient target identification and target area recommendation method based on physical constraint AI interpretation, enabling the inversion products to form a continuous processing chain of physical feature stack, weakly supervised probabilistic interpretation, reliability constraints, and target area ranking output.
[0008] To achieve the above-mentioned objectives, the technical solution provided by this invention includes: A SQUID magnetic gradient target identification and target area recommendation method based on physical constraint AI interpretation includes: Acquire SQUID magnetic gradient measurement data and its physical inversion results in the mining area; Construct a pixel-level physical feature stack based on the physical inversion results; Based on the pixel-level physical feature stack, pseudo-label samples are generated that do not use the ground truth labels of the current test area, and the class probability model is trained using the pseudo-label samples. The category probability model yields the category probability, entropy ambiguity, and in-distribution reliability of underground objects. Target area recommendation results are generated based on the underground object category probability, the entropy ambiguity, and the distribution reliability.
[0009] Preferably, the SQUID magnetic gradient measurement data of the mining area includes five independent magnetic gradient tensor components, a magnetic induction component, a coordinate grid, and an effective data mask. Constructing the pixel-level physical feature stack includes reconstructing the tensor component Gzz based on the five independent magnetic gradient tensor components and calculating the gradient invariant.
[0010] Preferably, the physical inversion results include multiple factors from the equivalent magnetic source, source depth, magnetic moment, source location uncertainty, source density, and posterior source density obtained by sparse physical dictionary inversion or Group-SPD inversion.
[0011] Preferably, the pixel-level physical feature stack includes multiple items from the following: co-gridized observation invariants, physical fitting invariants, residual invariants, source density, shallow source density, depth proxy, posterior source density, texture features, geological prior map, and effective data mask.
[0012] Preferably, generating the pseudo-label samples includes generating background scores, mineralization-related magnetic proxy scores, non-value magnetic body scores, and shallow interference object scores based on the pixel-level physical feature stack.
[0013] Preferably, generating the pseudo-label samples further includes extraction according to a preset quantile threshold, category winner criteria, background conditions, and a training sample quantity limit. The extracted objects include the background score, the mineralization-related magnetic proxy score, the non-value magnetic body score, and the shallow interference object score.
[0014] Preferably, the category probability model outputs the category probability of the underground object using the pseudo-label sample as the training target. The category probability of the underground object includes mineralization-related magnetic proxy probability, non-value or interference object probability, and background probability. The mineralization-related magnetic proxy probability includes at least one of gold ore magnetic proxy probability and copper ore magnetic proxy probability.
[0015] Preferably, obtaining the entropy ambiguity and the in-distribution reliability includes calculating the entropy ambiguity based on the underground object category probability, and determining the in-distribution reliability based on the pixel features and the category center distance or category distribution distance of the pseudo-label sample.
[0016] Preferably, obtaining the underground object category probability includes, when the reliability within the distribution does not reach a preset reliability condition, mixing and correcting the underground object category probability with the conservative distributions corresponding to the background, non-value magnetic bodies, or shallow interference objects.
[0017] Preferably, generating the target area recommendation results includes sorting based on the probability of valuable targets, the probability of non-valued or interfering targets, the entropy ambiguity, the reliability within the distribution, the area of connected regions, local peaks, and region splitting rules. The output includes the category, coordinates, priority score, and recommendation reason.
[0018] Beneficial effects 1. This invention unifies SQUID magnetic gradient measurement data and physical inversion results into a pixel-level physical feature stack, so that the input of the class probability model is no longer limited to ordinary image texture, but includes physical constraint information such as observation invariants, fitting invariants, residual invariants, source density, depth proxy and posterior source density, thereby improving the interpretability and verifiability of target recognition results.
[0019] 2. This invention generates pseudo-label samples based on physical features, geological priors, and rule scoring. It does not use the current survey area's true labels for model training, which can reduce the risk of confusion between training and validation boundaries and make the source of pseudo-labels traceable to specific physical features and scoring conditions.
[0020] 3. After outputting the category probability, this invention further calculates the entropy ambiguity and in-distribution reliability, and provides hints or conservative corrections for areas with obvious category competition, out-of-distribution or low reliability, so that the recognition result not only gives the target category probability, but also reflects the judgment stability and interpretation risk.
[0021] 4. In the target area recommendation stage, this invention integrates the probability of valuable targets, the probability of non-value or interference, the entropy ambiguity, the reliability within the distribution, the area of connected regions, and local peak information, avoiding the delineation of target areas solely based on abnormal intensity. It can output target area recommendation results that include category, coordinates, priority score, and recommendation reasons, which facilitates subsequent manual review, supplementary testing planning, and verification deployment. Attached Figure Description
[0022] Figure 1 This is a flowchart illustrating a preferred embodiment of the SQUID magnetic gradient target identification and target area recommendation method for mining areas based on physical constraint AI interpretation. Figure 2 This is a schematic diagram of the reliability gating and entropy ambiguity calculation process provided in a preferred embodiment of the present invention. Detailed Implementation
[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments. The listed embodiments are used to illustrate the implementation of the present invention. Specific parameters, model types, thresholds, and output fields can be adjusted according to the quality of mining area data, the completeness of geological data, and verification objectives, and should not be construed as limiting the scope of protection of the present invention.
[0024] The physical constraints referred to in this invention refer to the constraints imposed on model input, pseudo-label generation, and low-reliability area correction by SQUID magnetic gradient tensor relations, physical inversion products, effective data masks, geological priors, and physical rule scoring, thus limiting the judgment of the class probability model to physical sources and data validity. The AI interpretation refers to the class probability model outputting the underground object class probability, entropy ambiguity, and in-distribution reliability based on the aforementioned physical feature stack, and further transforming this output into verifiable target area recommendation results and reasons for the recommendation, rather than directly detecting mineral types or metal elements.
[0025] Example 1 like Figure 1 As shown, this embodiment provides a method for target identification and target area recommendation in mining areas based on SQUID magnetic gradient interpretation using physical constraints. The method takes SQUID magnetic gradient measurement data and its physical inversion results in the mining area as input. After processing with a pixel-level physical feature stack, pseudo-label samples, and a category probability model, it obtains the subsurface object category probability, entropy ambiguity, distribution reliability, and target area recommendation results. The subsurface object categories can include background, mineralization-related magnetic proxies, non-value magnetic bodies, and shallow interference objects; wherein, gold and copper magnetic proxies only represent magnetic or geological proxy responses related to the corresponding mineralization systems, and do not represent direct detection of gold or copper elements. "Non-value" refers to non-target magnetic responses relative to the current mineralization-related target.
[0026] In this embodiment, target identification and target area recommendation are within the same processing chain. The category probability, entropy ambiguity, and in-distribution reliability output from the target identification stage serve as inputs for the target area recommendation stage. The target area recommendation stage does not establish separate mineral type identification rules, nor does it consider borehole layout methods as a necessary step. Figure 1 Correspondingly, the method includes the following five steps.
[0027] Step S1: Obtain SQUID magnetic gradient measurement data and its physical inversion results for the mining area.
[0028] The SQUID magnetic gradient measurement data for the mining area includes five independent magnetic gradient tensor components, a magnetic induction component, a coordinate grid, and an effective data mask. The five independent magnetic gradient tensor components may include... , , , and The magnetic induction component can be The coordinate grid is used to determine the position of each grid point in the survey area plane, and the effective data mask is used to identify the effective observation area that can participate in subsequent calculations.
[0029] When measurement data contains random missing values, large blocks of missing values, strong local anomalies, or background trends, missing value imputation, detrending, and robust smoothing can be performed under the constraint of an effective data mask. Missing value imputation can use neighborhood interpolation or mask-based local regression; detrending can use low-order polynomial fitting of the background trend; and robust smoothing is used to reduce the impact of isolated noise points on subsequent invariants and feature stacks. The above preprocessing serves the purpose of measurement data acquisition and inversion result access, and does not constitute a necessary method step independent of step S1.
[0030] Based on the property that the magnetic gradient tensor is approximately traceless in passive observation space, the tensor components can be reconstructed using the following formula. .
[0031]
[0032] in, and These represent the two diagonal gradient components of the magnetic flux density along the plane coordinate direction; This represents the vertical diagonal gradient component obtained under unscented constraints. The above reconstruction enables the five independent magnetic gradient tensor components to form a complete symmetric tensor approximation.
[0033] In one specific implementation, the first Gradient invariants at each grid point It can be calculated using the following formula.
[0034]
[0035] in, , , , , and Indicates the first The corresponding magnetic gradient tensor components at each grid point; This represents the gradient invariant at this grid point. This invariant, characterized by the combined intensity of tensor components, helps to reduce the influence of coordinate orientation selection on anomaly intensity display.
[0036] The physical inversion results can be obtained by calling the inversion module internally within this method, or provided by preceding physical inversion software or an independent inversion system. The physical inversion results include multiple parameters from the equivalent magnetic source, source depth, magnetic moment, source location uncertainty, source density, and posterior source density obtained from sparse physical dictionary inversion or Group-SPD inversion. For this method, the physical inversion results serve as the input interface for subsequent target recognition; the complete inversion algorithm is not the sole source for target recognition and target area recommendation methods in this embodiment.
[0037] Step S2: Construct a pixel-level physical feature stack based on the physical inversion results.
[0038] Measurement data, gradient invariants, and physical inversion results are transformed to the same coordinate grid to form a pixel-level physical feature stack. This grid-based transformation can include spatial interpolation, mask alignment, resolution matching, and boundary clipping, ensuring a one-to-one correspondence between physical images, inversion source densities, and prior maps from different sources at the pixel level.
[0039] The pixel-level physical feature stack may include multiple items from the following: observation invariants, physical fitting invariants, residual invariants, source density, shallow source density, depth proxy, posterior source density, texture features, geological prior maps, and effective data masks. Observation invariants can be obtained by logarithmic compression and robust normalization of the gradient invariants obtained in step S1; physical fitting invariants can be calculated by forward modeling from equivalent magnetic sources; residual invariants represent the difference between observation invariants and physical fitting invariants, used to indicate model mismatch, shallow source interference, or unexplained structures.
[0040] Source density represents the spatial concentration of equivalent magnetic sources on a planar grid. Shallow source density emphasizes the contribution of near-surface sources to anomalies. Depth proxies can be used to distinguish shallow disturbance objects from deeper mineralization-related magnetic proxies. Posterior source density represents source location uncertainty or multiple solution support. Texture features can be formed by local gradients, anisotropy, patch boundaries, or local statistics. Effective data masks are used to avoid forcing definitive interpretations in invalid regions.
[0041] In one specific implementation, the first The first pixel Each feature can be normalized using a robust metric:
[0042] in, Indicates the first The first pixel One original or derived feature; This represents the corresponding normalized feature; Indicates the first The median of each feature within the valid data area; Indicates the absolute deviation of the median; This represents a very small positive number used to avoid division by zero. Using robust scaling helps reduce the impact of local strong outliers and isolated values on the characteristic distribution.
[0043] Geological prior maps can be derived from at least one of the following: geological mapping, geochemical anomalies, remote sensing alteration, tectonic lines, lithological boundaries, electrical or electromagnetic data, historical engineering data, and expert rules. If a project has geochemical anomalies, remote sensing alteration, polarizability, resistivity, electromagnetic conductors, tectonic lines, lithological boundaries, or existing borehole data, these data can be converted into geological prior maps and used as optional channels in the pixel-level physical feature stack. Different data can be normalized according to spatial resolution and confidence level and then superimposed, or they can be input as multiple prior channels for subsequent scoring. If a project lacks a geological prior map, the feature stack can set the corresponding channel to a null value marker, a low confidence constant, or a weak prior map generated by physical features, and reduce the weight of that prior channel in subsequent scoring.
[0044] Step S3: Generate pseudo-label samples that do not use the ground truth labels of the current test area based on the pixel-level physical feature stack, and train the class probability model using the pseudo-label samples.
[0045] This step first generates background scores, mineralization-related magnetic proxy scores, non-value magnetic body scores, and shallow interference object scores based on a pixel-level physical feature stack. Then, pseudo-label samples are extracted according to preset quantile thresholds, category winner conditions, background conditions, and training sample quantity limits. These pseudo-label samples are then used as training targets to train the category probability model. The pseudo-label samples are generated from physical features, geological priors, and rule-based scoring, and do not use the current survey area's ground truth labels. Historical ground truths, synthetic ground truths, or field verification results can be used for post-training evaluation, parameter sensitivity analysis, or expert review, but are not included in the training sample generation process of this step.
[0046] The mineralization-related magnetic proxy score can be generated by setting separate magnetic proxy scores for gold and copper deposits, and then combining these two scores to form the mineralization-related magnetic proxy score. (Gold deposit magnetic proxy score) It can be determined by the following formula.
[0047]
[0048] Copper Ore Magnetic Agent Rating It can be determined by the following formula.
[0049]
[0050] in, and They represent the first Pixel-level magnetic proxy priors for gold mines and copper mines; Indicates source density; Indicates the physical fit invariants; Represents texture features; Indicates the posterior source density; Indicates shallow source density; Represents the residual invariants; to , to These are pre-set or determined scores based on project data. The terms "gold ore" and "copper ore" in the above scores are magnetic proxy categories and do not represent direct measurements of the metallic elements themselves.
[0051] Magnetic agent rating related to mineralization The larger value between the magnetic proxy score of the gold ore and the magnetic proxy score of the copper ore can be taken.
[0052]
[0053] Non-value magnetic body rating It can be determined by the following formula.
[0054]
[0055] Scoring of shallow interference objects It can be determined by the following formula.
[0056]
[0057] Background rating It can be determined by the following formula.
[0058]
[0059] in, Indicates deep proxy; to , to , to These are scoring coefficients, either preset or determined based on project data. These coefficients are used to reflect the contribution of various characteristics to the corresponding scores and are not required to be fixed values across all mining areas.
[0060] The positive and negative contributions of the above scores correspond to the physical meaning of the SQUID magnetic gradient inversion results. For mineralization-related magnetic surrogate scores, source density and physical fit invariants indicate the presence of anomaly responses at that location that can be explained by equivalent magnetic sources. Posterior source density indicates that the response is supported in uncertain samples. Geological priors and texture features are used to indicate the spatial background or tectonic morphology related to mineralization; these factors contribute positively. When shallow source density is high, the anomaly is more likely to originate from near-surface disturbances or engineering remnants. When residual invariants are high, the anomaly is inconsistent with the current physical model or is affected by noise or shallow sources; therefore, both are deductible items in the mineralization-related magnetic surrogate scores.
[0061] For non-valued magnetic bodies, higher source density, observational invariants, texture features, and residual invariants indicate the presence of magnetic responses or unexplained anomalous structures. However, the score is increased when gold and copper magnetic proxy priors and shallow source density do not support their inclusion as mineralization-related targets. For shallow disturbance objects, shallow source density, residual invariants, and observational invariants provide evidence of strong near-surface responses, while depth proxy and mineralization-related priors are used to suppress misclassification of deeper or geologically supported anomalies as shallow disturbances. Background scoring uses complementary terms of source density, observational invariants, and physical fitting invariants, subtracting mineralization-related priors, to select background pixels with weak anomalous responses and weak mineralization support.
[0062] The scoring coefficients can be determined by expert rules, historical mining area playback, pseudo-label distribution calibration, or quantile statistics of project data. Different mineralization-related magnetic proxy categories can be set for different deposit types. For example, in porphyry copper deposits or skarn copper deposits, the magnetic proxy priors for copper deposits, deep source density, and tectonic boundary relationships can be emphasized; in orogenic gold deposits or epithermal gold deposits, the magnetic proxy priors for gold deposits, linear tectonic textures, and specific depth proxy combinations can be emphasized. If a certain type of prior data is missing, a scoring rule based solely on physical inversion features is used, or the weight of that prior channel in the scoring is reduced.
[0063] When extracting pseudo-label samples, for each non-background category, pixels within the valid data area whose scores fall within a preset high quartile and meet the category winner criteria can be selected as candidate samples. The category winner criteria can be expressed as the target category score being higher than other category scores, or the difference between the target category score and the second-highest category score reaching a preset margin. For the background category, pixels with low source density, low observation invariants, low physical fitting invariants, weak mineralization priors, and high background scores can be selected as background pseudo-labels.
[0064] The training sample size limit is used to prevent an excessive number of samples from a single class from biasing the class probability model. This limit can be set as an upper limit on the number of samples per class, an upper limit on the number of samples per connected region, or an upper limit formed by spatial block sampling. If the number of candidate pseudo-labels for a certain class is insufficient, the confidence weight of the corresponding class in this round of training can be reduced, or the class can be reserved for manual review or subsequent retesting.
[0065] The category probability model takes feature vectors from the pixel-level physical feature stack as input and the aforementioned pseudo-label categories as training targets, outputting the probability that each pixel belongs to each underground object category. The category probability model can be ExtraTrees, Random Forest, Gradient Boosting Tree, Probabilistic Calibration Support Vector Machine, Shallow Neural Network, or other models capable of outputting category probabilities. For models that cannot directly output normalized category probabilities, probability calibration mapping, category voting ratio normalization, or soft maximization output transformation can be used to obtain the probabilities of each category, making the probability distributions of underground object categories at the same pixel comparable. Preferably, a probability model that can provide feature importance or local explanatory information is used to verify the model output against the physical feature sources.
[0066] The output of the categorical probability model can be denoted as: ,in Indicates pixel index, This represents the subsurface object category index. The subsurface object category probability may include mineralization-related magnetic proxy probability, non-value or interference object probability, and background probability. The mineralization-related magnetic proxy probability may further include at least one of gold ore magnetic proxy probability and copper ore magnetic proxy probability; the non-value or interference object probability may include at least one of non-value magnetic body probability and shallow interference object probability. The background probability, as a subsurface object category probability parallel to the mineralization-related magnetic proxy probability and the non-value or interference object probability, is used to represent the probability that the corresponding pixel belongs to the background category.
[0067] Step S4: Obtain the underground object category probability, entropy ambiguity, and distribution reliability from the category probability model.
[0068] like Figure 2As shown, after the category probability model outputs the underground object category probability for all grid pixels, it can further calculate entropy ambiguity and in-distribution reliability. Entropy ambiguity represents the degree of competition among multiple categories at the same pixel, while in-distribution reliability represents whether the pixel feature is close to the pseudo-label training distribution. When the in-distribution reliability does not meet the preset reliability condition, a conservative distribution mixing correction can be applied to the underground object category probability; this correction is a preferred processing method when obtaining the underground object category probability and is not a main process step independent of this step. It should be understood that the conservative distribution refers to a category probability distribution that tends to favor background, non-value magnetic objects, or shallow interference objects at low-reliability pixels, used to reduce the risk of model over-interpretation.
[0069] In one specific implementation, the first Entropy ambiguity per pixel It can be calculated using the following formula.
[0070]
[0071] in, Indicates the first The pixel belongs to the first The probability of each category; Indicates the number of categories; This represents the smallest positive number used to avoid taking the logarithm of zero; This represents the entropy ambiguity after normalization. The higher the value, the closer the competition between multiple categories, and the higher the interpretive ambiguity of that pixel.
[0072] Intra-distribution reliability can be determined based on the distance between pixel features and the class centers or class distribution distances of the pseudo-label samples. In one specific implementation, the class centers of each class of pseudo-label samples are first calculated in the standardized feature space. and discrete scale . No. Normalized distance of pixels It can be calculated using the following formula.
[0073]
[0074] in, Indicates the first A standardized feature vector of 1 pixel; Indicates the first The pseudo-label category center for each category; Indicates the first Discrete scale of each category; Indicates Euclidean distance; Indicates the first The normalized distance from each pixel to the distribution of the pseudo-label category closest to the target pixel.
[0075] No. Intra-pixel distribution reliability It can be obtained from the normalized distance using the following formula.
[0076]
[0077] in, The closer This indicates that the pixel is closer to the pseudo-label training distribution in the feature space; The closer This indicates that the pixel may be located in an out-of-distribution region, a missing measurement boundary region, or a region where the category is difficult to interpret. In other embodiments, the in-distribution reliability can also be determined based on the category distribution distance, local density, Mahalanobis distance, or a calibrated out-of-distribution fraction.
[0078] When the reliability of a pixel's distribution does not meet the preset reliability condition, directly using the class probability model output may lead to over-interpretation of missing data, noise, or out-of-distribution areas. Therefore, this embodiment uses a hybrid correction of the pixel's underground object class probability with a conservative distribution. The conservative distribution can correspond to the background, non-value magnetic objects, or shallow interference objects, and can also be selected based on the pixel's shallow source density, residual invariants, and effective data mask.
[0079] In one specific implementation, the corrected class probability It can be determined by the following formula.
[0080]
[0081] in, This represents the original class probability output by the class probability model; In a conservative distribution, the th The pixel belongs to the first The probability of each category; This represents the mixing coefficient. The mixing coefficient can increase with increasing in-distribution reliability.
[0082] in, This represents a preset reliability scale. The preset reliability scale can be determined based on the normalized distance distribution from the pseudo-label samples to their class centers, historical mining area playback results, validation set calibration results, or project data quantile statistics; in the absence of a validation set, a representative quantile in the pseudo-label training distribution can be selected as the reliability scale, rather than being limited to a fixed value. If it is higher, then near The corrected probability mainly retains the output of the class probability model; if Lower, then The reduced probability means the probability is more influenced by the conservative distribution. For pixels outside the invalid data mask, a conservative distribution with a higher background probability can be directly output, or they can be marked as blind spots or candidate regions for supplementary testing.
[0083] The construction of a conservative distribution can be determined based on project data. For example, when shallow source density and residual invariants are high, a conservative distribution can increase the probability of shallow interfering objects; when source density, observational invariants, and geological priors are all weak, a conservative distribution can increase the background probability; when source density is high but geological priors are weak and depth proxies do not support target mineralization, a conservative distribution can increase the probability of non-value magnetic bodies. This treatment ensures that results in low-reliability areas are primarily interpreted conservatively, reducing the risk of model over-interpretation.
[0084] Step S5: Generate target area recommendation results based on the underground object category probability, the entropy ambiguity, and the distribution reliability.
[0085] This step transforms the underground object category probabilities, entropy ambiguity, and in-distribution reliability obtained in step S4 into target area candidates, rankings, and recommendation rationales. The target area recommendation results are based on the target identification results and do not use rules unrelated to target identification to re-delineate the target area.
[0086] In one specific implementation, the probability of the value target It can be determined by the magnetic proxy probability of gold ore and the magnetic proxy probability of copper ore.
[0087]
[0088] in, Indicates the first The probability of magnetic proxy for gold mines per pixel; Indicates the first The magnetic proxy probability of copper ore per pixel. This target probability follows the aforementioned meaning of magnetic proxy and does not represent a direct determination of the ore grade or metallic elements.
[0089] Non-value or interference probability It can be determined by the following formula. This probability is used to evaluate non-value or interfering factors in target candidate screening and ranking, excluding the background probability that is parallel to it.
[0090]
[0091] in, Indicates the probability of a non-valued magnetic object; This represents the probability of shallow interference objects. This represents the background probability, which can be used as a co-category probability for output explanation, invalid area marking, or manual review prompts, but is not included in the non-value or interference probability.
[0092] When forming candidate masks for target areas, pixels that meet the following conditions can be selected: the probability of a valuable target reaches a preset value condition; the probability of a valuable target is higher than the probability of a non-value or interference target by a preset margin; the entropy ambiguity is lower than a preset ambiguity condition; and the reliability within the distribution reaches a preset reliability condition. The above value condition, margin, ambiguity condition, and reliability condition can be determined based on the project verification stage, the required number of target areas, and risk preference.
[0093] After performing connectivity analysis on the candidate mask for the target region, several candidate connected regions can be obtained. If the area of a certain connected region is too large, or if there are multiple local peaks in the region and the peak spacing meets the preset splitting condition, the connected region is split according to the local peaks and the region area to avoid a single large patch swallowing multiple independent target regions. Specifically, local peaks of the probability or priority score of valuable targets can be extracted in the connected region first, and local peaks that meet the preset peak intensity condition and minimum spacing condition can be used as region seeds; then, the connected regions are assigned according to the distance from the pixel to each region seed, the watershed boundary along the probability surface, or the attribution relationship with the nearest peak; when the saddle probability between two seeds does not meet the preset descent condition, they can also be kept as the same candidate region. The center coordinates, area, number of pixels, average probability of valuable targets, average probability of non-value or interference targets, average entropy ambiguity, average reliability within the distribution, and dominant category of each candidate target region can be calculated.
[0094] In one specific implementation, the first Priority score of each target candidate region It can be determined by the following formula.
[0095]
[0096] in, Indicates the first The average probability of valuable targets within each candidate target region; This represents the average entropy ambiguity within the region. This represents the average reliability within the distribution area. This represents the average probability of non-value or interference within the region. , , and This is a preset weight or scaling factor. This score gives higher rankings to target regions that are high-value, low-ambiguity, high-reliability, and low-non-value interference.
[0097] The target area recommendation results can include category, coordinates, priority score, and recommendation reason. The recommendation reason can be generated jointly by the dominant category, probability of valuable targets, probability of non-value or interference, entropy ambiguity, reliability within the distribution, area of connected regions, and local peak information. For example, the recommendation reason can be stated as "This area has a high probability of mineralization-related magnetic proxy, low entropy ambiguity, reliability within the distribution meets the preset reliability conditions, and the area of connected regions reaches the preset candidate target area scale." For areas with high value but high entropy or low reliability, the recommendation reason can be marked as "Recommendation for supplementary testing or manual verification."
[0098] The target area recommendation results can also be used as input for subsequent borehole candidate point generation, field reconnaissance, supplementary survey planning, or expert review systems. Each target area entry may include a sort number, dominant category, center coordinates, area, number of pixels, magnetic surrogate probability for gold mines, magnetic surrogate probability for copper mines, non-value or interference probability, entropy ambiguity, reliability within the distribution, priority score, and recommendation reason. Subsequent systems can use high-value, low-entropy, and high-reliability target areas as high-confidence verification objects, high-value but high-entropy areas as interpretation and discrimination objects, high-priority but low-reliability areas near missing survey areas as supplementary survey objects, and areas with high non-value or interference probability as exclusion or avoidance objects. This interface does not change the target identification and target area recommendation mainline of the present invention method, but rather uses the output of this method as a structured input for subsequent deployment decisions.
[0099] The above embodiments illustrate one implementation of the present invention in the interpretation of SQUID magnetic gradient inversion results. For different survey areas, different deposit types, or different data completeness, specific features, categories, thresholds, models, and ranking coefficients can be adjusted without changing the physical feature stack, pseudo-labels, category probabilities, entropy ambiguity, distribution reliability, and target area recommendation mainline.
[0100] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.
Claims
1. A method for target identification and target area recommendation in SQUID magnetic gradient mining areas based on AI interpretation of physical constraints, characterized in that, include: Acquire SQUID magnetic gradient measurement data and its physical inversion results in the mining area; Construct a pixel-level physical feature stack based on the physical inversion results; Based on the pixel-level physical feature stack, pseudo-label samples are generated that do not use the ground truth labels of the current test area, and the class probability model is trained using the pseudo-label samples. The category probability model yields the category probability, entropy ambiguity, and in-distribution reliability of underground objects. Target area recommendation results are generated based on the underground object category probability, the entropy ambiguity, and the distribution reliability.
2. The SQUID magnetic gradient target identification and target area recommendation method for mining areas based on physical constraint AI interpretation as described in claim 1, characterized in that, The SQUID magnetic gradient measurement data of the mining area includes five independent magnetic gradient tensor components, magnetic induction components, coordinate grids, and effective data masks. Constructing the pixel-level physical feature stack includes reconstructing the tensor component Gzz based on the five independent magnetic gradient tensor components and calculating the gradient invariant.
3. The SQUID magnetic gradient target identification and target area recommendation method for mining areas based on physical constraint AI interpretation as described in claim 2, characterized in that, The physical inversion results include multiple parameters from the equivalent magnetic source, source depth, magnetic moment, source location uncertainty, source density, and posterior source density obtained by sparse physical dictionary inversion or Group-SPD inversion.
4. The SQUID magnetic gradient target identification and target area recommendation method for mining areas based on physical constraint AI interpretation according to claim 3, characterized in that, The pixel-level physical feature stack includes multiple items from the following: observation invariants, physical fitting invariants, residual invariants, source density, shallow source density, depth proxy, posterior source density, texture features, geological prior map, and effective data mask.
5. The SQUID magnetic gradient target identification and target area recommendation method for mining areas based on physical constraint AI interpretation according to claim 4, characterized in that, The generation of the pseudo-label samples includes generating background scores, mineralization-related magnetic proxy scores, non-value magnetic body scores, and shallow interference object scores based on the pixel-level physical feature stack.
6. The SQUID magnetic gradient target identification and target area recommendation method for mining areas based on physical constraint AI interpretation according to claim 5, characterized in that, Generating the pseudo-label samples also includes extracting the pseudo-label samples from the background score, the mineralization-related magnetic proxy score, the non-value magnetic body score, and the shallow interference object score according to a preset quantile threshold, category winner condition, background condition, and training sample quantity limit.
7. The SQUID magnetic gradient target identification and target area recommendation method for mining areas based on physical constraint AI interpretation according to claim 6, characterized in that, The category probability model outputs the category probability of the underground object using the pseudo-label sample as the training target. The category probability of the underground object includes mineralization-related magnetic proxy probability, non-value or interference object probability, and background probability. The mineralization-related magnetic proxy probability includes at least one of gold ore magnetic proxy probability and copper ore magnetic proxy probability.
8. The SQUID magnetic gradient target identification and target area recommendation method based on physical constraint AI interpretation according to claim 7, characterized in that, Obtaining the entropy ambiguity and the in-distribution reliability includes calculating the entropy ambiguity based on the underground object category probability, and determining the in-distribution reliability based on the pixel features and the category center distance or category distribution distance of the pseudo-label sample.
9. The SQUID magnetic gradient target identification and target area recommendation method for mining areas based on physical constraint AI interpretation according to claim 8, characterized in that, The process of obtaining the probability of the underground object category includes, when the reliability within the distribution does not meet the preset reliability condition, mixing and correcting the probability of the underground object category with the conservative distributions corresponding to the background, non-value magnetic bodies, or shallow interference objects.
10. The SQUID magnetic gradient target identification and target area recommendation method for mining areas based on physical constraint AI interpretation according to claim 9, characterized in that, The generation of the target area recommendation result includes outputting a target area recommendation result containing category, coordinates, priority score and recommendation reason based on the probability of valuable targets, the probability of non-value or interference, the entropy ambiguity, the reliability within the distribution, the area of connected regions, local peaks and region splitting rules.
