Three-dimensional induced polarization data inversion and target identification method

Abstract

The application provides a kind of three-dimensional induced polarization inversion and target identification method based on deep learning, comprising: preprocessing three-dimensional induced polarization observation data collected by multiple sources to obtain preprocessed induced polarization data;Based on non-structural multi-scale grid, a three-dimensional induced polarization forward and inversion model is constructed, the preprocessed induced polarization data is input into the three-dimensional induced polarization forward and inversion model, and iterative solution is carried out combined with grid self-adaptive encryption and joint inversion objective function to generate three-dimensional electrical property characteristic data;Apply mixed regularization constraint to three-dimensional electrical property characteristic data to generate high-resolution three-dimensional electrical property data volume;The mixed regularization constraint contains smoothing constraint and total variation constraint, and the weight proportion of the smoothing constraint and the total variation constraint is dynamically adjusted according to the detection depth;Input the high-resolution three-dimensional electrical property data volume into the deep learning identification model to output the target identification result.The application effectively eliminates the false anomaly of complex fault structure, greatly improves the accuracy and efficiency of deep mineral exploration, and is suitable for fine exploration scene of deep deposit.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration data processing technology, and in particular to a method for three-dimensional induced polarization data inversion and target identification. Background Technology

[0002] In deep mineral resource exploration, induced polarization (IP) is an important method for detecting concealed metallic ore bodies. However, as the exploration depth increases to the kilometer level, 3D IPC inversion faces severe challenges. On the one hand, signal attenuation at depth leads to a decrease in the signal-to-noise ratio. Traditional single-regularization inversion struggles to balance the boundary sharpness of shallow anomalies with the model continuity of weak signal regions in deep depths, often resulting in overly smooth shallow layers or divergent deep layers, leading to low accuracy in delineating the boundaries of concealed deep targets. On the other hand, in complex geological structures, the joint inversion of multiple physical properties such as resistivity and polarizability is highly susceptible to the influence of prior structures such as fault zones. Without effective spatial structural constraints, the inversion results often produce false transition zone artifacts across faults, severely interfering with the judgment of the true ore body morphology. Therefore, there is an urgent need for a 3D IPC data inversion and target identification method that can adaptively balance the resolution between shallow and deep layers and effectively suppress structural artifacts. Summary of the Invention

[0003] Therefore, it is necessary to provide a method for three-dimensional induced polarization data inversion and target recognition to address the aforementioned technical problems.

[0004] A method for three-dimensional induced polarization data inversion and target identification includes the following steps:

[0005] Preprocessing is performed on the three-dimensional induced polarization (IP) observation data acquired from multiple sources to obtain preprocessed IPO data;

[0006] A three-dimensional induced polarization (IP) forward and inverse model is constructed based on an unstructured multi-scale grid. The preprocessed IPI data is input into the three-dimensional IPI forward and inverse model. The model is then iteratively solved by combining adaptive grid densification and a joint inverse objective function to generate three-dimensional electrical characteristic data.

[0007] A hybrid regularization constraint is applied to the three-dimensional electrical feature data to generate a high-resolution three-dimensional electrical data volume; the hybrid regularization constraint includes a smoothing constraint and a total variation constraint, and the weight ratio of the smoothing constraint and the total variation constraint is dynamically adjusted according to the detection depth;

[0008] The high-resolution three-dimensional electrical data volume is input into a deep learning recognition model, and the target recognition result is output.

[0009] In one embodiment, the multi-source acquired three-dimensional induced polarization (IP) observation data is preprocessed to obtain preprocessed IPO data, including:

[0010] Convert multi-source acquired three-dimensional induced polarization observation data into a unified standard spatial multidimensional array format;

[0011] Kriging interpolation based on the correlation of the variation function of spatial neighborhood electrode data is used to fill in scattered missing values:

[0012]

[0013] in, Indicates lag distance The semivariogram value at that location, This represents the spatial distance vector between two observation electrode points. Indicates the distance is The total number of observation points Indicates position Observed values ​​of polarization or resistivity at the excitation electrode. Indicates position The observed values ​​of polarization or resistivity at the excitation electrode.

[0014] In one embodiment, the iterative solution is performed by combining grid adaptive encryption and joint inversion objective function, including:

[0015] During the iterative solution process, the subdivision of the unstructured multi-scale mesh is triggered based on the local element residual indicator factor;

[0016] The local element residual indicator factor is calculated by weighting the volume residual term inside the element and the flux jump surface residual term on the element boundary:

[0017]

[0018] in, Representation unit Local error estimator on Representation unit Feature size, For the residual terms within the cell, Representing an edge Length, Representing an edge Boundary residuals on Representation unit The boundary.

[0019] In one embodiment, the formula for calculating the joint inversion objective function is as follows:

[0020]

[0021] in, This represents the total objective function value. This represents the model of the subsurface medium physical property parameters to be inverted. This represents the synthetic data generated by the forward simulation operator. This represents actual observation data. Represents the data weight matrix. Represents the regularization coefficient. Represents the model weight matrix. This represents the reference model.

[0022] In one embodiment, the smoothing constraint is an L2 norm-based constraint, and the total variation constraint is an L1 norm-based constraint.

[0023] The step of dynamically adjusting the weight ratio of the smoothing constraint and the total variation constraint based on the detection depth includes:

[0024] In the shallow region, the weight ratio of the total variation constraint relative to the smoothness constraint is increased;

[0025] In deeper regions, the weighting of the smoothing constraint relative to the total variation constraint is increased:

[0026]

[0027] in, Indicates about model parameters The regularization objective function; and These represent the weight coefficients of the smoothing constraint and the total variation constraint, respectively. Indicates the volume of the inversion space; Represents a weighted matrix; Represents the gradient operator; Represents the reference model; This represents a small regularization constant used to avoid numerical instability when the gradient is zero.

[0028] In one embodiment, the joint inversion objective function further includes a multi-property parameter space cross gradient constraint term, wherein the multi-property parameters include resistivity and polarizability;

[0029] The multi-property parameter spatial cross-gradient constraint term is constructed by calculating the square of the modulus of the cross product of the spatial gradient vector field of resistivity and the spatial gradient vector field of polarizability:

[0030]

[0031] in, As a gradient correlation index, The resistivity model parameter distribution obtained by inversion, For the distribution of parameters in the polarizability model, The spatial gradient vector field of the resistivity model. The spatial gradient vector field of the polarizability model. This represents the vector cross product operation.

[0032] This forces the resistivity and polarizability to maintain a parallel or antiparallel distribution in three-dimensional space.

[0033] In one embodiment, the deep learning recognition model is a three-dimensional convolutional neural network. The step of inputting the high-resolution three-dimensional electrical data volume into the deep learning recognition model and outputting the target recognition result includes:

[0034] The high-resolution three-dimensional electrical data volume is input into the three-dimensional convolutional neural network;

[0035] A three-dimensional spatial self-attention mechanism module is embedded in the deep semantic extraction layer of the three-dimensional convolutional neural network. Through the interactive weighting of the query matrix, key matrix, and value matrix, the deep target region with strong induced electrostatic response gradient is given an increased activation weight.

[0036]

[0037] in, Indicates attention weights, Represents the query matrix. Represents the key matrix. Represents a value matrix, The dimension of the key vector. This represents the matrix transpose operation. This represents the normalization function.

[0038] In one embodiment, the three-dimensional convolutional neural network is trained in the following manner:

[0039] Backpropagation training is performed using a composite function that combines Dice coefficient loss and focus cross-entropy loss.

[0040] The imbalance between positive and negative samples between the target anomaly and the background region is addressed by adjusting the weights of easy and difficult samples using a category balance factor and a focus parameter.

[0041]

[0042] in, This represents the total loss value; and These represent the weight coefficients of the Dice loss term and the cross-entropy loss term, respectively; Indicates the first The predicted probability value of each sample; Indicates the first The true label value of each sample; Indicates the category balance factor; This indicates the focus parameter.

[0043] In one embodiment, it further includes:

[0044] Obtain prior data on the three-dimensional spatial orientation of geological fault zones;

[0045] When constructing the unstructured multi-scale mesh, the unstructured multi-scale mesh is forced to break at the physical edge of the geological fault zone or extend along the fault strike, so as to cooperate with the spatial cross gradient constraint term of the multi-property parameter to eliminate artifacts caused by smoothing across faults.

[0046] A three-dimensional induced polarization (IP) data inversion and target recognition system is provided to implement the three-dimensional IPS data inversion and target recognition method described above, comprising:

[0047] The data preprocessing module is used to preprocess the three-dimensional induced polarization observation data acquired from multiple sources to obtain preprocessed induced polarization data;

[0048] The forward and inverse calculation module is used to construct a three-dimensional induced polarization (IP) forward and inverse model based on an unstructured multi-scale grid. The preprocessed IPI data is input into the three-dimensional IPI forward and inverse model, and the model is iteratively solved by combining adaptive grid densification and joint inverse objective function to generate three-dimensional electrical characteristic data.

[0049] The regularization constraint module is used to apply hybrid regularization constraints to the three-dimensional electrical feature data to generate a high-resolution three-dimensional electrical data volume; the hybrid regularization constraints include smoothing constraints and total variation constraints, and the weight ratio of the smoothing constraints and the total variation constraints is dynamically adjusted according to the detection depth;

[0050] The target recognition module is used to input the high-resolution three-dimensional electrical data volume into the deep learning recognition model and output the target recognition result.

[0051] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: The present invention constructs a depth-adaptive hybrid regularization constraint mechanism, dynamically adjusts the weight ratio of smoothing constraint and total variation constraint according to the detection depth, solves the inherent contradiction between shallow boundary ambiguity and deep noise amplification in deep induced polarization inversion at the physical level, and realizes stable reconstruction of high-resolution three-dimensional electrical data volume in the full depth range.

[0052] This invention introduces a multi-property parameter spatial cross gradient constraint term and combines it with an unstructured meshing strategy guided by prior geological structure to force the resistivity and polarizability models to maintain the consistency of gradient direction in three-dimensional space. It eliminates the false transition zone artifacts caused by cross-fault smoothing in complex tectonic areas from the two dimensions of mathematical constraints and geometric discretization, and significantly improves the geological credibility of the boundary characterization of hidden target bodies.

[0053] This invention embeds a three-dimensional spatial self-attention mechanism and uses a joint loss function for training during the deep learning recognition stage. This effectively overcomes the difficulties in extracting features of weak signals in deep areas and the extreme imbalance between positive and negative samples. It enables intelligent and accurate delineation of the boundaries of concealed ore bodies from massive amounts of three-dimensional electrical data, significantly reducing the subjectivity and uncertainty of manual interpretation. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating a deep learning-based three-dimensional induced polarization inversion and target recognition method in one embodiment.

[0055] Figure 2 This is a schematic diagram of the structure of a deep learning-based three-dimensional induced polarization inversion and target recognition system in one embodiment. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0057] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of this specification should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of this specification do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word covers the element or object listed following the word and its equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0058] In one embodiment, such as Figure 1 As shown, a method for three-dimensional induced polarization data inversion and target recognition is provided, including the following steps:

[0059] Step S101: Preprocess the three-dimensional induced polarization observation data acquired from multiple sources to obtain preprocessed induced polarization data.

[0060] Specifically, the three-dimensional induced polarization response data collected from multiple sources in both well-to-surface and inter-well environments are standardized in format, including data sampling frequency, amplitude unit, and timestamp. Based on spatial correlation, outlier removal, missing value filling, and strong noise suppression are performed to achieve dynamic data quality assessment and obtain high-quality preprocessed induced polarization data.

[0061] Step S102: Construct a three-dimensional induced polarization forward and inverse model based on an unstructured multi-scale grid. Input the preprocessed induced polarization data into the three-dimensional induced polarization forward and inverse model, and iteratively solve the model by combining grid adaptive densification and joint inversion objective function to generate three-dimensional electrical characteristic data.

[0062] Specifically, based on the physical properties of deep mineral deposits and the topology of the observation system, a three-dimensional induced polarization forward and inverse physical model is constructed using an unstructured multi-scale tetrahedral Delaunay subdivision grid. By combining local grid adaptive refinement technology with structural coupling joint inversion objective function, high-precision forward and inverse calculations are completed through iterative solution, and independent three-dimensional apparent resistivity and three-dimensional polarizability physical components are obtained through decoupling.

[0063] Step S103: Apply hybrid regularization constraints to the three-dimensional electrical feature data to generate a high-resolution three-dimensional electrical data volume; wherein, the hybrid regularization constraints include smoothing constraints and total variation constraints, and the weight ratio of the smoothing constraints and the total variation constraints is dynamically adjusted according to the detection depth.

[0064] Specifically, a hybrid regularization algorithm is introduced that combines the L1 norm (sum of the absolute values ​​of all elements of a vector) total variation constraint with the L2 norm (square root of the sum of the squares of all elements of a vector) smoothing constraint. The weight ratio of the two types of constraints is dynamically adjusted according to the detection depth to suppress the ill-conditioned multiple solutions and deep dispersion phenomenon in the inversion iteration process, and generate a high-precision three-dimensional induced polarization feature data volume with a grid resolution better than 10m that truly reflects the underground electrical distribution.

[0065] Step S104: Input the high-resolution three-dimensional electrical data volume into the deep learning recognition model and output the target recognition result.

[0066] Specifically, multi-dimensional implicit representations of three-dimensional induced polarization feature data volumes are extracted, a standardized three-dimensional feature tensor dataset is constructed, and the dataset is input into a three-dimensional convolutional neural network (3D-CNN) deep learning model with embedded spatiotemporal attention mechanism. Through model feature extraction and nonlinear mapping training, intelligent extraction of the boundary of the deep ore-rich area and quantitative estimation of ore content are achieved.

[0067] In this embodiment, by constructing a hybrid regularization model that includes smoothness constraints and total variation constraints, and dynamically adjusting the weight ratio of the two according to the detection depth, the problem that fixed regularization parameters in traditional inversion cannot adapt to signal characteristics at different depths is solved. At the same time, by combining unstructured multi-scale grid adaptive encryption technology, the numerical simulation accuracy of key areas is improved while ensuring computational efficiency, providing a high-quality data foundation for subsequent deep learning recognition.

[0068] Based on this, the multi-source acquired three-dimensional induced polarization (IP) observation data is preprocessed to obtain preprocessed IPE data, including:

[0069] Convert multi-source acquired three-dimensional induced polarization observation data into a unified standard spatial multidimensional array format;

[0070] Kriging interpolation based on the correlation of the variation function of spatial neighborhood electrode data is used to fill in scattered missing values:

[0071]

[0072] in, Indicates lag distance The semivariogram value at that location, This represents the spatial distance vector between two observation electrode points. Indicates the distance is The total number of observation points Indicates position Observed values ​​of polarization or resistivity at the excitation electrode. Indicates position The observed values ​​of polarization or resistivity at the excitation electrode.

[0073] Specifically, the time-domain and frequency-domain induced polarization waveform data output by different acquisition devices are converted into a unified standard spatial multidimensional array format, with a unified sampling frequency of 24ksps and a baseband signal amplitude unit of mV / V. Microsecond-level timestamp alignment is achieved through the IEEE 1588 protocol.

[0074] Ordinary kriging interpolation based on the correlation of the variogram function of spatial neighborhood electrode data is used to fill in missing values ​​in scattered data. The formula for calculating the variogram function is as follows:

[0075] ;

[0076] in, Indicates lag distance The semivariogram value at that location, This represents the spatial distance vector between two observation electrode points. Indicates the distance is The total number of observation points Indicates position Observed values ​​of polarization or resistivity at the excitation electrode. Indicates position The observed values ​​of polarization or resistivity at the excitation electrode.

[0077] For abnormal continuous data segments with a polarization mean square error greater than 15% and an apparent resistivity spatial jump gradient exceeding 50 Ω·m / m, a dynamic sliding window mid-range filtering method with 10 to 50 sampling points combined with segmented marking of local anomaly factors is used to hard-remove them.

[0078] For anomalous frequency bands with channel noise floor greater than 10nV / rt(Hz)@1Hz and receiver dynamic range less than 108dB@32ksps, a weighted matrix attenuation strategy is adopted to apply extremely small confidence weight coefficients in the range of 0.01~0.1 to achieve strong noise suppression.

[0079] By introducing a spatial variation function to quantify the spatial correlation of electrode data, and using Kriging interpolation to perform unbiased optimal estimation of missing data, this method can better preserve the spatial statistical characteristics of the original data compared to traditional linear interpolation, providing complete data input that conforms to geostatistical laws for subsequent inversion.

[0080] Based on this, an iterative solution is performed by combining grid adaptive encryption and joint inversion objective function, including:

[0081] During the iterative solution process, the subdivision of the unstructured multi-scale mesh is triggered based on the local element residual indicator factor;

[0082] The local element residual indicator factor is calculated by weighting the volume residual term inside the element and the flux jump surface residual term on the element boundary:

[0083]

[0084] in, Representation unit Local error estimator on Representation unit Feature size, For the residual terms within the cell, Representing an edge Length, Representing an edge Boundary residuals on Representation unit The boundary.

[0085] The formula for calculating the joint inversion objective function is as follows:

[0086]

[0087] in, This represents the total objective function value. This represents the model of the subsurface medium physical property parameters to be inverted. This represents the synthetic data generated by the forward simulation operator. This represents actual observation data. Represents the data weight matrix. Represents the regularization coefficient. Represents the model weight matrix. This represents the reference model.

[0088] Specifically, the mesh construction and inversion iteration process includes: establishing a multi-parameter coupled three-dimensional electromagnetic background model of the ore body covering conductivity sigma and polarizability eta; using unstructured tetrahedral Delaunay partitioning to discretize the response space; the minimum control side length of the mesh in the core detection area is 5~10m; and the external boundary mesh far from the field source is adaptively increased and diffused with a scaling factor of 1.15~1.25.

[0089] Based on the posterior error estimation theory, adaptive mesh refinement is triggered by the local element residual indicator factor eta_K. The formula for calculating the residual indicator factor is:

[0090]

[0091] in, Represents a tetrahedral element. Representation unit circumcircle diameter, Representation unit Feature size, For the residual terms within the cell, Represents grid edges Length, Representing an edge Boundary residuals on Representation unit The boundary.

[0092] The mesh is subdivided and refined in regions where eta_K is greater than 10% of the global average error; a structural synchronous coupling joint inversion overall objective function is constructed, and the iterative update calculation formula is as follows:

[0093]

[0094] in, This represents the total objective function value. This represents the model of the subsurface medium physical property parameters to be inverted. This represents the synthetic data generated by the forward simulation operator. This represents actual observation data. Represents the data weight matrix. Represents the regularization coefficient. Represents the model weight matrix. This represents the reference model.

[0095] The inversion iteration convergence is completed when the relative convergence error is less than 0.05 or the number of iterations reaches the upper limit of 200.

[0096] By constructing a local element residual indicator factor through posterior error estimation theory, and comprehensively considering the field value fitting error and boundary flux continuity error within the element, precise triggering of mesh refinement is achieved. Local refinement is performed only in areas where the numerical solution gradient changes drastically or the fitting residual is large, avoiding the waste of computational resources caused by global uniform refinement, and significantly improving the computational efficiency and accuracy of 3D forward and inverse modeling.

[0097] A standard objective function framework combining data fitting terms and model regularization terms was constructed. Noise interference was adaptively suppressed by the data weight matrix, and prior information constraints were introduced by the model weight matrix, ensuring the numerical stability and convergence of the inversion iteration process.

[0098] Based on this, the smoothing constraint is an L2 norm-based constraint, and the total variation constraint is an L1 norm-based constraint;

[0099] The step of dynamically adjusting the weight ratio of the smoothing constraint and the total variation constraint based on the detection depth includes:

[0100] In the shallow region, the weight ratio of the total variation constraint relative to the smoothness constraint is increased;

[0101] In deeper regions, the weighting of the smoothing constraint relative to the total variation constraint is increased:

[0102]

[0103] in, Indicates about model parameters The regularization objective function; and These represent the weight coefficients of the smoothing constraint and the total variation constraint, respectively. Indicates the volume of the inversion space; Represents a weighted matrix; Represents the gradient operator; Represents the reference model; This represents a small regularization constant used to avoid numerical instability when the gradient is zero.

[0104] Specifically, the hybrid regularization algorithm works as follows: based on the traditional L2 norm Tijinov smoothing regularization, an L1 norm three-dimensional total variation (TV) regularization term is introduced to construct a hybrid regularization constraint function, the calculation formula of which is:

[0105]

[0106] in, Indicates about model parameters The regularization objective function, This represents the overall smoothness weighting coefficient of the model, with a value ranging from 0.01 to 0.1. The weighting coefficient represents the sharpness of abrupt changes in physical property boundaries. Represents the depth-weighted matrix. Represents the inversion space volume. It represents the smallest positive number in the range of 10^(-6) to 10^(-8), used to ensure the continuity of the derivative and avoid numerical instability.

[0107] The weight ratio is dynamically adjusted based on the detection depth: for shallow to mid-layer regions with a detection depth Z < 1000m, / The ratio is increased to over 10 to enhance the clarity of the physical boundaries of anomalies; it can probe deep regions with depths of 1000m ≤ Z ≤ 3000m. / The ratio is reduced to below 0.1 to maintain the continuity of deep geological structures and to balance the stability of deep data with the resolution of local anomalies.

[0108] To address the physical characteristic of deep induced polarization signals attenuating with depth, a depth-adaptive weight adjustment strategy was designed: in shallow regions with strong signals and high resolution, the L1 norm total variation constraint weight is increased to sharpen the ore body boundary; in deep regions with weak signals and noise dominance, the L2 norm smoothing constraint weight is increased to maintain the continuity of the geological structure and suppress noise amplification. This mechanism fundamentally solves the contradiction that single regularization cannot simultaneously take into account the imaging quality of both shallow and deep layers.

[0109] Based on this, the joint inversion objective function also includes a multi-property parameter space cross gradient constraint term, wherein the multi-property parameters include resistivity and polarizability;

[0110] The multi-property parameter spatial cross-gradient constraint term is constructed by calculating the square of the modulus of the cross product of the spatial gradient vector field of resistivity and the spatial gradient vector field of polarizability:

[0111]

[0112] in, As a gradient correlation index, The resistivity model parameter distribution obtained by inversion, For the distribution of parameters in the polarizability model, The spatial gradient vector field of the resistivity model. The spatial gradient vector field of the polarizability model. This represents the vector cross product operation.

[0113] This forces the resistivity and polarizability to maintain a parallel or antiparallel distribution in three-dimensional space.

[0114] Specifically, a cross-gradient penalty constraint term is added to the objective function, and the calculation formula is as follows:

[0115]

[0116] in, As a gradient correlation index, The resistivity model parameter distribution obtained by inversion, For the distribution of parameters in the polarizability model, , These are the three-dimensional spatial gradient vectors of the two types of parameters. By penalizing the cross gradient values, the resistivity and polarizability models are forced to maintain a parallel or antiparallel distribution in three-dimensional space, eliminating false anomaly artifacts caused by fault smoothing interpolation and restoring the true physical morphology of the ore body.

[0117] By utilizing the principle of cross-gradient structural coupling, without forcing a linear correlation between resistivity and polarizability, only the consistency of their spatial gradient directions is constrained. This constraint can effectively utilize the structural information of high signal-to-noise ratio parameters (such as resistivity) to guide the inversion of low signal-to-noise ratio parameters (such as polarizability). Especially in complex structural regions, it can force the alignment of multiple physical property boundaries and completely eliminate the false transition zone artifacts caused by the forced smoothing of the mesh across faults.

[0118] Based on this, the deep learning recognition model is a three-dimensional convolutional neural network. The process of inputting the high-resolution three-dimensional electrical data volume into the deep learning recognition model and outputting the target recognition result includes:

[0119] The high-resolution three-dimensional electrical data volume is input into the three-dimensional convolutional neural network;

[0120] A three-dimensional spatial self-attention mechanism module is embedded in the deep semantic extraction layer of the three-dimensional convolutional neural network. Through the interactive weighting of the query matrix, key matrix, and value matrix, the deep target region with strong induced electrostatic response gradient is given an increased activation weight.

[0121]

[0122] in, Indicates attention weights, Represents the query matrix. Represents the key matrix. Represents a value matrix, The dimension of the key vector. This represents the matrix transpose operation. This represents the normalization function.

[0123] Specifically, three core representation parameters—3D gradient magnitude, spatial connectivity, and apparent polarizability voxel distribution density—are extracted from the high-resolution feature data volume to construct a standardized 3D feature tensor dataset with dimensions D*H*W*C. D represents the depth dimension, corresponding to the underground exploration depth direction (Z-axis, 0~3000m), H represents the longitudinal spatial dimension of the survey area (Y-axis), W represents the lateral spatial dimension of the survey area (X-axis), and C represents the number of feature channels. A 3D-CNN architecture with a downsampling encoder, an upsampling decoder, and cross-layer skip connections is adopted. Geological borehole physical property ground truth data is used as the supervision label, and the training and validation sets are divided in an 8:2 ratio. A 3D spatial self-attention mechanism is embedded in the deep semantic extraction layer of the encoder. The attention weight calculation formula is as follows:

[0124]

[0125] in, Represents the query matrix. Represents the key matrix. Represents a value matrix, The dimension of the key vector is used to automatically increase the model activation weights in deep, strong gradient target regions through an attention mechanism.

[0126] Embedding a spatial self-attention mechanism in a 3D convolutional neural network enables the model to adaptively focus on potential mineralized areas with significant changes in induced polarization response gradients, suppressing redundant feature interference from large-area background rocks. Through interactive weighting of Query-Key-Value (the three core matrices of the attention mechanism), the network's ability to extract weak but key anomalous features from deep areas is enhanced, improving the sensitivity and accuracy of target recognition.

[0127] Based on this, the three-dimensional convolutional neural network is trained in the following way:

[0128] Backpropagation training is performed using a composite function that combines Dice coefficient loss and focus cross-entropy loss.

[0129] The imbalance between positive and negative samples between the target anomaly and the background region is addressed by adjusting the weights of easy and difficult samples using a category balance factor and a focus parameter.

[0130]

[0131] in, This represents the total loss value; and These represent the weight coefficients of the Dice loss term and the cross-entropy loss term, respectively; Indicates the first The predicted probability value of each sample; Indicates the first The true label value of each sample; Indicates the category balance factor; This indicates the focus parameter.

[0132] Specifically, to address the volume sample imbalance problem between deep small ore bodies and surrounding rock at a 1:100 scale, a composite loss function combining Dice coefficient (a measure of the similarity between the predicted and actual labeled regions, ranging from 0 to 1) loss and focal cross-entropy loss is used for model training. The total loss is calculated as follows:

[0133]

[0134] in, This represents the total loss value; and These represent the weight coefficients of the Dice loss term and the cross-entropy loss term, respectively; Indicates the first The predicted probability value of each sample; Indicates the first The true label value of each sample; Indicates the category balance factor; This indicates the focus parameter.

[0135] After the model converges, it performs forward inference on the input data volume, outputs a three-dimensional probability heat map, filters voxels with a confidence level greater than 85%, quantitatively calculates the ore-bearing percentage of the anomaly, and completes the intelligent identification of the ore body.

[0136] In addition to this, it also includes:

[0137] Obtain prior data on the three-dimensional spatial orientation of geological fault zones;

[0138] When constructing the unstructured multi-scale mesh, the unstructured multi-scale mesh is forced to break at the physical edge of the geological fault zone or extend along the fault strike, so as to cooperate with the spatial cross gradient constraint term of the multi-property parameter to eliminate artifacts caused by smoothing across faults.

[0139] Specifically, a multi-dimensional prior structure cross-gradient constraint mechanism is added: extract the three-dimensional spatial orientation prior mapping data of the geological fault zone in the survey area, construct a rigid prior structure model, and force the grid to break along the edge of the fault zone or extend along the fault orientation when dividing the unstructured grid.

[0140] Integrating prior geological structural information into the grid partitioning process ensures that the grid topology matches the actual underground geological interface. This geometric constraint, together with the aforementioned cross-gradient physical constraint, creates a synergistic effect, avoiding artificial cross-layer smoothing in numerical simulation from the source of discretization, and further guaranteeing the geological rationality of the inversion results for complex tectonic zones.

[0141] This invention solves the inherent contradiction between shallow boundary ambiguity and deep noise amplification in deep induced polarization inversion at the physical level by constructing a depth-adaptive hybrid regularization constraint mechanism and dynamically adjusting the weight ratio of smoothing constraint and total variation constraint according to the detection depth, thereby achieving stable reconstruction of high-resolution three-dimensional electrical data volume across the entire depth range.

[0142] This invention introduces a multi-property parameter spatial cross gradient constraint term and combines it with an unstructured meshing strategy guided by prior geological structure to force the resistivity and polarizability models to maintain the consistency of gradient direction in three-dimensional space. It eliminates the false transition zone artifacts caused by cross-fault smoothing in complex tectonic areas from the two dimensions of mathematical constraints and geometric discretization, and significantly improves the geological credibility of the boundary characterization of hidden target bodies.

[0143] This invention embeds a three-dimensional spatial self-attention mechanism and uses a joint loss function for training during the deep learning recognition stage. This effectively overcomes the difficulties in extracting features of weak signals in deep areas and the extreme imbalance between positive and negative samples. It enables intelligent and accurate delineation of the boundaries of concealed ore bodies from massive amounts of three-dimensional electrical data, significantly reducing the subjectivity and uncertainty of manual interpretation.

[0144] It should be noted that the method of this embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this embodiment, and the multiple devices will interact with each other to complete the method described.

[0145] It should be noted that the above description describes some embodiments of the present invention. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than that shown in the above embodiments and still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0146] The technical solutions of this invention will be clearly and completely described below with reference to the embodiments thereof. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0147] Example 1

[0148] This embodiment is applied to the "exploration and assessment" scenario in deep old metal mines. The exploration target is a disseminated blind ore body at a depth of 2000-3000m. The surrounding rock properties in this area are complex, and weak mineralization signals are easily masked by background noise. Traditional methods have a high rate of missed detection and false detection. The specific implementation steps are as follows:

[0149] Step S1: Multi-source data standardization preprocessing. Multi-source induced polarization (IP) data from the well-to-surface high-power detection system and the inter-well high-resolution acquisition system are collected and uniformly mapped to a three-dimensional Gauss-Kruger absolute coordinate system. The standardized sampling frequency is 24 ksps, and the signal unit is mV / V, achieving microsecond-level timestamp alignment. Ordinary Kriging interpolation is used to fill in scattered missing values. Abnormal data segments with a polarization mean square error >15% and resistivity gradient >50 Ω·m / m are removed through threshold filtering. Signal frequency bands with excessive noise and insufficient dynamic range are subject to minimal weighting coefficients to suppress strong noise, resulting in clean, standardized preprocessed data.

[0150] Step S2: High-precision forward and inverse calculation using unstructured multi-scale grids. A coupled electromagnetic background model of multiple physical parameters in the mining area is constructed, using unstructured tetrahedral grids. The minimum side length of the grid in the core exploration area is set to 8m, while the outer area is adaptively diffused and increased at a ratio of 1.2. During the inversion iteration, the grid accuracy is judged in real time by the element residual indicator factor, and abnormal areas with excessive errors are automatically subdivided and densified. Based on the structurally coupled joint inversion objective function, the preprocessed nonlinear conjugate gradient method is used for iterative solution until the convergence error is less than 0.05, completing the decoupled output of the three-dimensional apparent resistivity and polarizability components.

[0151] Step S3: Reconstruction of high-resolution data volume using hybrid regularization. The regularization weights are adaptively adjusted based on the exploration depth: the total variation constraint weights are amplified in areas shallower than 1000m to enhance the sharpness of orebody boundaries; the smoothing constraint weights are amplified in areas deeper than 1000-3000m to maintain the continuity of the stratigraphic structure. Through iterative optimization using L1 and L2 hybrid regularization, the problems of inversion ambiguity and deep data divergence are thoroughly suppressed, generating a three-dimensional high-precision electrical feature data volume with a resolution better than 10m.

[0152] Step S4: Deep Learning Intelligent Target Recognition. Gradient magnitude, spatial connectivity, and voxel distribution density features of the data volume are extracted to construct a 3D feature tensor dataset, which is then divided into training and validation sets in an 8:2 ratio. A 3D-CNN model with embedded spatiotemporal attention mechanism is used, combined with a composite loss function, to complete model training. This automatically focuses on deep, subtle anomaly target areas, effectively solving the problem of imbalanced samples in small ore bodies. After model convergence, a 3D probability heatmap is output, and voxels with over 85% confidence are selected to automatically delineate ore body boundaries and quantitatively estimate ore content. The ore body recognition accuracy can reach 95.2%.

[0153] The entire process is automated and visualized. Through a microservice architecture, the system schedules tasks across the entire process, automatically generating multi-angle X, Y, and Z slices of the 3D stratigraphy. Ore body identification results are output in VTK format. The system automatically calculates inversion time, data fitting error, target intersection-over-union ratio, and other indicators, generating a standardized exploration assessment report to provide precise data support for drilling project deployment. In this embodiment, the overall inversion calculation time is reduced by 42% compared to traditional methods, perfectly adapting to the needs of deep, detailed exploration in old mines.

[0154] Example 2

[0155] This embodiment is applied to a complex geological exploration scenario involving intersecting fault zones and water-rich strata. In this area, signals are prone to phase distortion, and traditional methods can easily misidentify water-bearing faults as ore bodies. The specific implementation process is optimized by adding structural constraints based on the first embodiment.

[0156] In step S2, the mesh construction and inversion process, the existing three-dimensional prior mapping data of geological fault zones in the survey area are imported to construct a rigid prior structural model. This forces the unstructured mesh to break at the physical edges of the fault zones and extend along the fault strike. A cross-gradient penalty constraint term is embedded in the joint inversion objective function to force the spatial gradient directions of the resistivity and polarizability models to remain coupled and consistent, completely eliminating false transition zone artifacts caused by mesh smoothing across faults and accurately restoring the true shape of the ore body dissected by faults.

[0157] In step S4, during model training, a dedicated negative sample library is constructed to address the pseudo-anomaly characteristics of water-rich faults. Through positive and negative sample adversarial training, the model's ability to distinguish between real ore bodies and structural disturbances is enhanced, thereby improving the model's robustness.

[0158] After optimization in this embodiment, the IoU (Intersection over Union) ratio for identifying ore bodies in complex geological areas is stabilized above 0.88, which can accurately identify false anomalies caused by faults and groundwater, completely solving the problem of exploration misjudgment in complex geological scenarios and significantly reducing the cost of ineffective drilling.

[0159] Based on the same inventive concept, corresponding to any of the above embodiments, the present invention also provides a three-dimensional induced polarization data inversion and target recognition system.

[0160] refer to Figure 2 The three-dimensional induced polarization data inversion and target recognition system includes:

[0161] Data preprocessing module 201 is used to preprocess the three-dimensional induced polarization observation data acquired from multiple sources to obtain preprocessed induced polarization data;

[0162] The forward and inverse calculation module 202 is used to construct a three-dimensional induced polarization forward and inverse model based on an unstructured multi-scale grid. The preprocessed induced polarization data is input into the three-dimensional induced polarization forward and inverse model. The model is iteratively solved by combining grid adaptive densification and joint inverse objective function to generate three-dimensional electrical characteristic data.

[0163] The regularization constraint module 203 is used to apply hybrid regularization constraints to the three-dimensional electrical feature data to generate a high-resolution three-dimensional electrical data volume; the hybrid regularization constraints include smoothing constraints and total variation constraints, and the weight ratio of the smoothing constraints and the total variation constraints is dynamically adjusted according to the detection depth.

[0164] The target recognition module 204 is used to input the high-resolution three-dimensional electrical data volume into the deep learning recognition model and output the target recognition result.

[0165] For ease of description, the above system is described by dividing it into various modules based on their functions. Of course, in implementing this invention, the functions of each module can be implemented in one or more software and / or hardware components.

[0166] The system described in the above embodiments is used to implement a corresponding three-dimensional induced polarization data inversion and target recognition method in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0167] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of the invention as described above, which are not provided in the details for the sake of brevity.

[0168] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0169] While specific details have been set forth to describe exemplary embodiments of the invention, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive. Although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description.

[0170] The embodiments of this invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this invention should be included within the protection scope of this invention.

Claims

1. A method for three-dimensional induced polarization data inversion and target recognition, characterized in that, include: Preprocessing is performed on the three-dimensional induced polarization (IP) observation data acquired from multiple sources to obtain preprocessed IPO data; A three-dimensional induced polarization (IP) forward and inverse model is constructed based on an unstructured multi-scale grid. The preprocessed IPI data is input into the three-dimensional IPI forward and inverse model. The model is then iteratively solved by combining adaptive grid densification and a joint inverse objective function to generate three-dimensional electrical characteristic data. A hybrid regularization constraint is applied to the three-dimensional electrical feature data to generate a high-resolution three-dimensional electrical data volume; wherein, the hybrid regularization constraint includes a smoothing constraint and a total variation constraint, and the weight ratio of the smoothing constraint and the total variation constraint is dynamically adjusted according to the detection depth; The high-resolution three-dimensional electrical data volume is input into a deep learning recognition model, and the target recognition result is output.

2. The method for three-dimensional induced polarization data inversion and target recognition according to claim 1, characterized in that, The preprocessing of multi-source acquired three-dimensional induced polarization (IP) observation data to obtain preprocessed IPE data includes: Convert multi-source acquired three-dimensional induced polarization observation data into a unified standard spatial multidimensional array format; Kriging interpolation based on the correlation of the variation function of spatial neighborhood electrode data is used to fill in scattered missing values: in, Indicates lag distance The semivariogram value at that location, This represents the spatial distance vector between two observation electrode points. Indicates the distance is The total number of observation points Indicates position Observed values ​​of polarization or resistivity at the excitation electrode. Indicates position The observed values ​​of polarization or resistivity at the excitation electrode.

3. The method for three-dimensional induced polarization data inversion and target recognition according to claim 1, characterized in that, The iterative solution combining grid adaptive encryption and joint inversion objective function includes: During the iterative solution process, the subdivision of the unstructured multi-scale mesh is triggered based on the local element residual indicator factor; The local element residual indicator factor is calculated by weighting the volume residual term inside the element and the flux jump surface residual term on the element boundary: in, Representation unit Local error estimator on Representation unit Feature size, For the residual terms within the cell, Representing an edge Length, Representing an edge Boundary residuals on Representation unit The boundary.

4. The method for three-dimensional induced polarization data inversion and target recognition according to claim 1, characterized in that, The formula for calculating the joint inversion objective function is as follows: in, This represents the total objective function value. This represents the model of the subsurface medium physical property parameters to be inverted. This represents the synthetic data generated by the forward simulation operator. This represents actual observation data. Represents the data weight matrix. Represents the regularization coefficient. Represents the model weight matrix. This represents the reference model.

5. The method for three-dimensional induced polarization data inversion and target recognition according to claim 1, characterized in that, The smoothing constraint is an L2 norm-based constraint, and the total variation constraint is an L1 norm-based constraint. The step of dynamically adjusting the weight ratio of the smoothing constraint and the total variation constraint based on the detection depth includes: In the shallow region, the weight ratio of the total variation constraint relative to the smoothness constraint is increased; In deeper regions, the weighting of the smoothing constraint relative to the total variation constraint is increased: in, Indicates about model parameters The regularization objective function; and These represent the weight coefficients of the smoothing constraint and the total variation constraint, respectively. Indicates the volume of the inversion space; Represents a weighted matrix; Represents the gradient operator; Represents the reference model; This represents a small regularization constant used to avoid numerical instability when the gradient is zero.

6. The method for three-dimensional induced polarization data inversion and target recognition according to claim 4, characterized in that, The joint inversion objective function also includes a multi-property parameter space cross gradient constraint term, wherein the multi-property parameters include resistivity and polarizability; The multi-property parameter spatial cross-gradient constraint term is constructed by calculating the square of the modulus of the cross product of the spatial gradient vector field of resistivity and the spatial gradient vector field of polarizability: in, As a gradient correlation index, The resistivity model parameter distribution obtained by inversion, For the distribution of parameters in the polarizability model, The spatial gradient vector field of the resistivity model. The spatial gradient vector field of the polarizability model. This represents the vector cross product operation. This forces the resistivity and polarizability to maintain a parallel or antiparallel distribution in three-dimensional space.

7. The method for three-dimensional induced polarization data inversion and target recognition according to claim 1, characterized in that, The deep learning recognition model is a three-dimensional convolutional neural network. The process of inputting the high-resolution three-dimensional electrical data volume into the deep learning recognition model and outputting the target recognition result includes: The high-resolution three-dimensional electrical data volume is input into the three-dimensional convolutional neural network; A three-dimensional spatial self-attention mechanism module is embedded in the deep semantic extraction layer of the three-dimensional convolutional neural network. Through the interactive weighting of the query matrix, key matrix, and value matrix, the deep target region with strong induced electrostatic response gradient is given an increased activation weight. in, Indicates attention weights, Represents the query matrix. Represents the key matrix. Represents a value matrix, The dimension of the key vector. This represents the matrix transpose operation. This represents the normalization function.

8. The method for three-dimensional induced polarization data inversion and target recognition according to claim 7, characterized in that, The three-dimensional convolutional neural network was trained in the following way: Backpropagation training is performed using a composite function that combines Dice coefficient loss and focus cross-entropy loss. The imbalance between positive and negative samples between the target anomaly and the background region is addressed by adjusting the weights of easy and difficult samples using a category balance factor and a focus parameter. in, This represents the total loss value; and These represent the weight coefficients of the Dice loss term and the cross-entropy loss term, respectively; Indicates the first The predicted probability value of each sample; Indicates the first The true label value of each sample; Indicates the category balance factor; This indicates the focus parameter.

9. The three-dimensional induced polarization data inversion and target recognition method according to claim 6, characterized in that, Also includes: Obtain prior data on the three-dimensional spatial orientation of geological fault zones; When constructing the unstructured multi-scale mesh, the unstructured multi-scale mesh is forced to break at the physical edge of the geological fault zone or extend along the fault strike, so as to cooperate with the spatial cross gradient constraint term of the multi-property parameter to eliminate artifacts caused by smoothing across faults.

10. A three-dimensional induced polarization data inversion and target recognition system, characterized in that, A method for implementing a three-dimensional induced polarization data inversion and target recognition method as described in any one of claims 1-9 includes: The data preprocessing module is used to preprocess the three-dimensional induced polarization observation data acquired from multiple sources to obtain preprocessed induced polarization data; The forward and inverse calculation module is used to construct a three-dimensional induced polarization (IP) forward and inverse model based on an unstructured multi-scale grid. The preprocessed IPI data is input into the three-dimensional IPI forward and inverse model, and the model is iteratively solved by combining adaptive grid densification and joint inverse objective function to generate three-dimensional electrical characteristic data. The regularization constraint module is used to apply hybrid regularization constraints to the three-dimensional electrical feature data to generate a high-resolution three-dimensional electrical data volume; the hybrid regularization constraints include smoothing constraints and total variation constraints, and the weight ratio of the smoothing constraints and the total variation constraints is dynamically adjusted according to the detection depth; The target recognition module is used to input the high-resolution three-dimensional electrical data volume into the deep learning recognition model and output the target recognition result.

Images