Deep learning and phased array radar wave velocity inversion-based deep geological ore body identification method

Abstract

The application discloses a deep geomineral body identification method based on deep learning and phased array radar wave velocity inversion, relates to the technical field of deep geomineral body identification, and comprises the following steps: S1, a two-dimensional or three-dimensional scanning is performed on underground space by a phased array deep geomineral body detection radar system with a preset frequency being controlled, and data acquisition is performed; S2, the collected data is preprocessed, and the wave velocity and multidimensional signal features of electromagnetic waves are extracted; S3, the wave velocity is inverted to obtain the three-dimensional distribution of dielectric constants and the three-dimensional distribution of equivalent conductivities; S4, the features are input into a pre-trained deep neural network model, and a mineral body type label and a prediction confidence are output; S5, the spatial position, the depth and the mineral body type identification result are fused to generate a three-dimensional mineral body distribution model, and a comprehensive exploration report is output. The application realizes rapid, nondestructive and accurate identification and evaluation of various industrial mineral bodies, and solves the industry problems of a shallow detection depth, single mineral identification and dependence on artificial experience in the traditional method.

Description

Technical Field

[0001] This invention relates to the field of deep-earth mineral body identification technology, specifically to a method for deep-earth mineral body identification based on deep learning and phased array radar wave velocity inversion. Background Technology

[0002] Ground-penetrating radar (GPR) technology, based on phased arrays, has been attempted for shallow and deep geological surveys as a non-destructive testing method. However, traditional GPR systems and methods based on inverting dielectric constants from early signal characteristics suffer from the following technical limitations: insufficient detection depth (conventional GPR systems are limited by antenna power and sensitivity, with effective detection depths often less than 30 meters, making it difficult to meet the needs of deep-earth mineral exploration); limited identification capabilities (existing methods mostly focus on lithological differentiation or single mineral identification, lacking a systematic wave velocity-dielectric property mapping model covering major industrial minerals, thus failing to effectively distinguish between minerals such as coal (low dielectric constant) and metallic minerals (high dielectric constant); low inversion accuracy and reliability, and high ambiguity (traditional methods rely on simplified empirical formulas or manual interpretations, making it difficult to handle complex electromagnetic wave propagation effects in deep environments (such as loss, dispersion, and noise). More importantly, geophysical inversion itself is a typical "ill-posed problem," meaning that multiple different subsurface media models can fit the observed radar data equally well; this phenomenon is known as "ambiguity." For example, a large low-grade mineralization zone and a small high-grade rich ore body may produce similar radar responses, which are difficult to distinguish based on a single parameter (such as wave velocity), resulting in insufficient reliability of the inversion results and interpretation results that are highly dependent on expert experience and have great uncertainty.

[0003] In summary, phased array radar technology has been successfully applied in the field of airborne radar through electronic scanning and beam focusing. However, applying this technology to deep-earth exploration (up to 120 meters and above) in high-loss geological media and integrating deep learning technology to achieve intelligent identification of multiple mineral types is still a technological gap. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for identifying deep earth mineral bodies based on deep learning and phased array radar wave velocity inversion.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] This invention discloses a method for identifying deep-earth ore bodies based on deep learning and phased array radar wave velocity inversion, comprising the following steps:

[0007] S1. The phased array deep-earth detection radar system with a preset frequency is used to perform two-dimensional or three-dimensional scanning of underground space and collect data.

[0008] S2. Preprocess the data collected in step S1, extract the wave velocity of electromagnetic waves based on tomography, and extract multi-dimensional signal features, including carrier frequency amplitude, signal envelope amplitude, signal attenuation coefficient, and signal spectrum centroid and bandwidth.

[0009] S3. The wave velocity extracted in step S2 is inverted to obtain the three-dimensional distribution of the dielectric constant, and the three-dimensional distribution of the equivalent conductivity is inverted based on the signal attenuation characteristics.

[0010] S4. The three-dimensional distribution of dielectric constant, the three-dimensional distribution of equivalent conductivity, and multi-dimensional signal features are used as input features and input into a pre-trained deep neural network model. Then, the model outputs the ore body type label and prediction confidence.

[0011] S5. Integrate the spatial location, depth, and ore body type identification results to generate a three-dimensional ore body distribution model and output a comprehensive exploration report, which includes ore body type, spatial morphology, burial depth, and confidence level.

[0012] Preferably, step S2 specifically includes:

[0013] Suppose the underground medium is discretized into N grid cells, and M radar wave ray paths pass through this region, then the... Propagation time of a ray for: ;in, Indicates the first The path length of the ray traveling through the j-th grid. This represents the slowness of the j-th grid. Indicates the first Random errors in the travel time observation of a ray;

[0014] Write all the ray equations in matrix form, i.e. , where L is A path matrix of dimension s, where s is A slowness vector of dimension t, where t is The observation travel time vector of dimension;

[0015] The solution is obtained using regularized least squares, i.e. ,in Describing the L2 norm, Represents the regularization parameter. The smoothness constraint matrix is ​​used to obtain the three-dimensional distribution of the slowness vector s, which is then converted into the three-dimensional distribution of the wave velocity v.

[0016] Preferably, in step S3, the formula is used. The dielectric constant is obtained by inverting the wave velocity v extracted in step S2. The three-dimensional distribution, and the equivalent conductivity inverted based on the signal attenuation characteristics. The three-dimensional distribution.

[0017] Preferably, step S3 also includes setting the dielectric constant. Three-dimensional distribution and equivalent conductivity The three-dimensional distribution is used for full waveform inversion, specifically including:

[0018] Define the objective function , Where m represents the model parameter vector. Indicates the first Radar signals measured at each receiving point This represents the result of forward simulation calculations using Maxwell's equations, given m. Radar signals from each receiving point;

[0019] The full waveform inversion updates the model parameters through iterative optimization, i.e. ,in, Indicates the number of iterations. Indicates the step size. Describe the objective function exist gradient at, This represents the inverse of the Hessian matrix.

[0020] Preferably, the architecture of the deep neural network model described in step S4 includes a convolutional neural network (CNN) and a long short-term memory network (E-LSTM).

[0021] Preferably, step S4 includes the following steps:

[0022] S41. Construct a feature vector from the features of each spatial location. ,in The characteristic dimensions include the three-dimensional distribution of the dielectric constant at that location, the three-dimensional distribution of the equivalent conductivity, and multi-dimensional signal characteristics.

[0023] S42. Treat the three-dimensional spatial data volume as a multi-channel image and input it into a deep neural network model. First, perform convolution through the convolutional layer of a convolutional neural network (CNN) to obtain the... Feature map of layer : ,in, Indicates the sequence number of the convolutional layer. This represents the convolution operation. Indicates the first Feature maps input to the layer, Indicates the first The convolutional kernel weights of the layer, Indicates the first Layer bias, This represents the activation function; subsequently, through multiple convolutional and pooling operations of the CNN, spatial features from local to global are extracted. ;

[0024] S43. Input the spatial features extracted by the CNN into the Long Short-Term Memory (E-LSTM) network, where the E-LSTM network... The internal state at time t is updated as follows: in, This represents the activation value of the forget gate. The weight matrix represents the forget gate. express The input feature vector at time t, This represents the hidden state at time t-1. and Input feature vector at time step To splice, This represents the bias vector of the forget gate. It is the sigmoid activation function; This represents the activation value of the input gate. This represents the weight matrix of the input gate. This represents the bias vector of the input gate; express Cellular state at any given moment express The state of candidate cells at any given time. Indicates calculation The weight matrix, Indicates calculation The bias vector, express t Cell state at time -1; This indicates the activation value of the output gate. express Output at any moment This represents element-wise multiplication. This represents the weight matrix of the output gate. The output gate's bias vector is represented; the gating mechanism based on E-LSTM outputs context features. ;

[0025] S44. The deep neural network model also includes an online clustering enhancement module, which uses a clustering loss function... Enhance the model for the samples The clustering loss function for: ,in Representing contextual features The features after projection network, where B represents the batch size. Indicates the assignment to the sample The cluster center vector, Indicates sample pseudo-tags;

[0026] S45, Finally, context features The input is fed into a fully connected layer to calculate the final probability distribution of ore body types: Where K represents the total number of ore body types, and c represents the ore body type number whose probability is to be calculated. This represents the weight vector corresponding to c. This represents the bias term corresponding to c, and m represents the ore body category number used for summation. This represents the weight vector corresponding to m. represents the bias term of m, and y represents the predicted ore body category label. The category with the highest probability is taken as the identification result, and its corresponding maximum probability value is taken as the prediction confidence.

[0027] Preferably, before the input features of different dimensions are input into the pre-trained deep neural network model, step S4 further includes fusing them: ,in, Represents the time-domain eigenvector. Represents the frequency domain eigenvector. This represents vector concatenation. This indicates the Z-Score normalization operation. and These represent the weight parameters, This represents the fused feature vector.

[0028] The beneficial effects of this invention are:

[0029] 1) This application improves the effective detection depth of GPR to more than 120 meters through an optimized phased array radar system, breaking through the bottleneck of traditional GPR in deep earth exploration.

[0030] 2) This application introduces deep hybrid neural networks (CNN + E-LSTM) and online clustering enhancement mechanisms into ore body identification. The model can automatically learn complex and nonlinear geological features, greatly reducing the reliance on human experience. The identification accuracy and mineral type coverage (from coal mines to metal mines) are significantly better than traditional methods.

[0031] 3) This application has strong anti-interference ability. The deep learning model used in this application, combined with multi-dimensional features (such as CFA) extracted from the original signal, can effectively combat noise and uncertainty in the deep earth environment and improve the reliability of the results.

[0032] 4) This application realizes a fully automated process from data acquisition to intelligent identification. The combination of rapid scanning by phased array radar and rapid inference by AI improves the efficiency of large-scale deep-earth mineral resource screening by several orders of magnitude.

[0033] 5) Traditional methods suffer from multiple solutions due to their reliance on a single parameter. This invention integrates multiple physical parameters such as wave velocity, dielectric constant, and conductivity with the amplitude-frequency characteristics of the original signal to form high-dimensional complementary input features. The deep learning model learns the complex nonlinear correlations and joint distributions between these features, greatly compressing the solution space and significantly reducing the multiple solutions in the inversion and recognition results. Attached Figure Description

[0034] Figure 1 This is a schematic diagram illustrating the steps of an embodiment of the present invention;

[0035] Figure 2 This is a schematic diagram of the phased array deep-earth detection radar system and deep learning processing center according to an embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of the architecture of a deep hybrid neural network (CNN + E-LSTM) according to an embodiment of the present invention;

[0037] Figure 4 This is a visualization diagram of the three-dimensional ore body type identification results in an embodiment of the present invention. Detailed Implementation

[0038] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] This invention discloses a deep-earth ore body identification method based on deep learning and phased array radar wave velocity inversion. It aims to solve the problems of shallow detection depth, limited ore type identification, and reliance on manual experience in existing technologies. The invention provides a deep-earth ore body identification method with greater detection depth, higher identification accuracy, and a higher degree of automation. The steps of the method are illustrated in the diagram below. Figure 1 As shown in the diagram, the structure of the phased array deep-earth detection radar system and the deep learning processing center is as follows: Figure 2 As shown, this method utilizes a phased array deep-earth detection radar, combined with a deep learning model, to accurately and non-destructively identify various industrial mineral bodies at depths exceeding 120 meters by inverting the electromagnetic wave velocity and dielectric properties within the earth. Specifically, it includes the following steps:

[0040] S1. Using a high-power, low-frequency optimized phased array deep-earth exploration radar system, broadband electromagnetic waves are emitted into the ground, and the beam pointing and focusing of the phased array antenna are electronically controlled to perform two-dimensional or three-dimensional scanning of the underground space and collect data. This step is optimized for typical geological media (such as granite and limestone), and its effective detection depth can reach more than 120 meters.

[0041] S2. Preprocess the data collected in step S1 (such as filtering and gain recovery), extract the wave velocity of electromagnetic waves based on tomography, and extract multi-dimensional signal features, including carrier frequency amplitude, signal envelope amplitude, signal attenuation coefficient, and signal spectrum centroid and bandwidth.

[0042] S3. The wave velocity extracted in step S2 is inverted to obtain the three-dimensional distribution of the dielectric constant, and the three-dimensional distribution of the equivalent conductivity is inverted based on the signal attenuation characteristics.

[0043] S4. The three-dimensional distribution of dielectric constant, the three-dimensional distribution of equivalent conductivity, and multi-dimensional signal features are used as input features and input into a pre-trained deep neural network model. Then, the model outputs the ore body type label and prediction confidence.

[0044] S5. Integrate the spatial location, depth, and ore body type identification results to generate a three-dimensional ore body distribution model and output a comprehensive exploration report. The comprehensive exploration report includes ore body type, spatial morphology, burial depth, and confidence level. A visualization of the three-dimensional ore body type identification results is shown below. Figure 4 As shown in Table 1, typical electromagnetic parameter reference ranges are provided.

[0045] Table 1: Reference Range of Typical Electromagnetic Parameters

[0046]

[0047] Specifically, the extraction of electromagnetic wave velocity using tomography in step S2 is mathematically essentially the solution of a large system of sparse linear equations, including:

[0048] Suppose the underground medium is discretized into N grid cells, and M radar wave ray paths pass through this region, then the... Propagation time of a ray for: ;in, Indicates the first The path length of the ray traveling through the j-th grid. This represents the slowness of the j-th grid. Indicates the first Random errors (or noise terms) in the travel time observation of a ray;

[0049] Write all the ray equations in matrix form, i.e. , where L is A path matrix of dimension s, where s is A slowness vector of dimension t, where t is The observation travel time vector of dimension;

[0050] because These are usually mixed-determined or overdetermined problems, and are solved using regularized least squares. ,in Describing the L2 norm, This represents the regularization parameter, used to balance the data fitting term with the model smoothing constraint term. The smoothing constraint matrix (usually a first- or second-order difference operator) is used to obtain the three-dimensional distribution of the slowness vector s, which is then converted into the three-dimensional distribution of the wave velocity v. This is the data fitting term, which measures the error between the model's predicted time travel and the measured time travel. This is a model regularization term used to constrain the smoothness of the solution and avoid overfitting.

[0051] Specifically, in step S3, the formula is used. The dielectric constant is obtained by inverting the wave velocity v extracted in step S2. The three-dimensional distribution, and the equivalent conductivity inverted based on the signal attenuation characteristics. The three-dimensional distribution.

[0052] Specifically, step S3 also involves the dielectric constant. Three-dimensional distribution and equivalent conductivity The three-dimensional distribution is used for full waveform inversion, with the goal of achieving optimal matching between the simulated radar waveform and the measured waveform to obtain more accurate dielectric constant and conductivity. Specifically, this includes:

[0053] Define the objective function , , where m represents the model parameter vector (containing the dielectric constant and conductivity of each grid). Indicates the first Radar signals measured at each receiving point This represents the result of forward simulation calculations using Maxwell's equations, given m. Radar signals from each receiving point;

[0054] The full waveform inversion updates the model parameters through iterative optimization (such as the conjugate gradient method or the L-BFGS method), i.e. ,in, Indicates the number of iterations. Indicates the step size. Describe the objective function exist The gradient at a given point can be efficiently calculated using the adjoint state method. This represents the inverse of the Hessian matrix, used to guide the update direction.

[0055] Specifically, the architecture of the deep neural network model described in step S4 includes a Convolutional Neural Network (CNN) and a Long Short-Term Memory (E-LSTM) network. The schematic diagram of the deep hybrid neural network (CNN + E-LSTM) architecture is shown below. Figure 3 As shown, a convolutional neural network (CNN) is used to extract local spatial features, and an enhanced long short-term memory network (E-LSTM) is used to capture contextual dependencies across spatial dimensions to understand the continuity and overall morphology of geological structures.

[0056] Specifically, step S4 includes the following steps:

[0057] S41. Construct a feature vector from the features of each spatial location. ,in The characteristic dimensions include the three-dimensional distribution of the dielectric constant at that location, the three-dimensional distribution of the equivalent conductivity, and multi-dimensional signal characteristics.

[0058] S42. Treat the three-dimensional spatial data volume (i.e., the three-dimensional distribution of dielectric constant, the three-dimensional distribution of equivalent conductivity, and the specific digital representation of multi-dimensional signal features) as a multi-channel image, input it into a deep neural network model, and first perform convolution through the convolutional layer of a convolutional neural network (CNN) to obtain the... Feature map of layer : ,in, Indicates the sequence number of the convolutional layer. This represents the convolution operation. Indicates the first Feature maps input to the layer, Indicates the first The convolutional kernel weights of the layer, Indicates the first Layer bias, This represents the activation function; subsequently, through multiple convolutional and pooling operations of the CNN, spatial features from local to global are extracted. ;

[0059] S43. Input the spatial features extracted by the CNN into the Long Short-Term Memory (E-LSTM) network, where the E-LSTM network... The internal state at time t is updated as follows: in, This represents the activation value of the forget gate. The weight matrix represents the forget gate. express The input feature vector at time t, This represents the hidden state at time t-1. and Input feature vector at time step To splice, This represents the bias vector of the forget gate. It is the sigmoid activation function; This represents the activation value of the input gate. This represents the weight matrix of the input gate. This represents the bias vector of the input gate; express Cellular state at any given moment express The state of candidate cells at any given time. Indicates calculation The weight matrix, Indicates calculation The bias vector, express t Cell state at time -1; This indicates the activation value of the output gate. express Output at any moment This represents element-wise multiplication. This represents the weight matrix of the output gate. The output gate's bias vector is represented; the gating mechanism based on E-LSTM outputs context features. The E-LSTM gating mechanism effectively captures cross-dimensional contextual dependencies in spatial features.

[0060] S44. The deep neural network model also includes an online clustering enhancement module. Therefore, during training, the model can dynamically discover and adapt to potential, undefined mineralization patterns in the data, enhancing its robustness in identifying unknown or complex mineralization types. The online clustering enhancement module uses a clustering loss function... Enhance the model to improve its ability to discover new patterns for samples. The clustering loss function for: ,in Representing contextual features The features after projection network, where B represents the batch size. Indicates the assignment to the sample The cluster center vector, Indicates sample The pseudo-labels; this step learns features into a compact cluster structure, thus remaining sensitive to mineralization patterns that are not explicitly labeled in the training data;

[0061] S45, Finally, context features The input is fed into a fully connected layer to calculate the final probability distribution of ore body types: Where K represents the total number of ore body types, and c represents the ore body type number whose probability is to be calculated. This represents the weight vector corresponding to c. This represents the bias term corresponding to c, and m represents the ore body category number used for summation. This represents the weight vector corresponding to m. represents the bias term of m, and y represents the predicted ore body category label. The category with the highest probability is taken as the identification result, and its corresponding maximum probability value is taken as the prediction confidence.

[0062] Specifically, before input features of different dimensions are fed into the pre-trained deep neural network model, step S4 also includes fusing them: ,in, Represents the time-domain eigenvector. Represents the frequency domain eigenvector. This represents vector concatenation. This indicates the Z-Score normalization operation. and These represent the weight parameters, enabling the model to focus on the most important feature sources based on different geological environments. This represents the fused feature vector.

[0063] For example, identifying lead-zinc ore bodies in polymetallic metallogenic areas:

[0064] Exploration background: A polymetallic metallogenic belt exists in a certain mountainous area. Traditional geophysical exploration methods show low resistivity anomalies, but cannot distinguish the mineral types.

[0065] Equipment and data acquisition: A phased array deep-earth detection radar with a center frequency of 50 MHz is used to collect three-dimensional data along the survey line, with a target depth of more than 120 meters.

[0066] Data processing and recognition:

[0067] Inversion yields the dielectric constant ( The three-dimensional data volume shows that there is a [missing information] at a depth of 95-110 meters. =15±2 anomalies.

[0068] Inverted conductivity ( The result shows that the abnormal body has an σ value of 5 ± 2 S / m.

[0069] The carrier frequency amplitude (CFA) extracted from the original signal exhibits a significantly high value in this region.

[0070] The aforementioned multi-dimensional features are input into a pre-trained deep hybrid neural network model.

[0071] Results and Validation: The model identified the anomaly as a lead-zinc ore body with a confidence level of 92%. Drilling confirmed this, revealing a thick lead-zinc ore layer at a depth of 102 meters, which highly matched the model prediction.

[0072] For example, identifying coal seams and their roof and floor in sedimentary basins.

[0073] Exploration Background: A coal resource area needs to accurately delineate the burial depth and thickness of coal seams to provide a design basis for mining.

[0074] Equipment and data acquisition: A phased array radar system with a center frequency of 25 MHz is used to obtain a greater penetration depth.

[0075] Data processing and recognition:

[0076] The wave velocity inversion profile clearly shows a low-velocity layer (wave velocity of approximately 0.13 m / ns) at a depth of 60-65 meters.

[0077] Dielectric constant inversion shows that the layer =4.5, significantly lower than the surrounding rock ( ≈7-9).

[0078] The conductivity is shown as low ( <0.01 S / m).

[0079] Deep learning models integrate these features to accurately identify coal seams and simultaneously delineate their roof (sandstone) and floor (mudstone).

[0080] Results and Validation: The 3D coal seam structure diagram output by the model shows an average coal seam thickness of 5 meters. The subsequent three validation boreholes all successfully encountered coal, with a burial depth error of less than 1 meter.

[0081] For example, distinguishing copper and tungsten deposits in skarn-type deposits.

[0082] Exploration background: In a certain skarn-type deposit, copper and tungsten ore coexist, but the beneficiation processes are different, so they need to be distinguished in advance.

[0083] Equipment and data acquisition: A phased array radar with a center frequency of 100 MHz is used to obtain higher resolution.

[0084] Data processing and recognition:

[0085] Within the same mineralized structure, inversion revealed two adjacent anomalies: anomaly A... =12, =50 S / m; Anomalous body B =14, =10 S / m.

[0086] Although their dielectric constants are similar, their conductivity and the "centroid of the spectrum" characteristics extracted from the spectrum are significantly different.

[0087] Based on these subtle differences, the deep learning model identifies anomaly A as copper ore and anomaly B as tungsten ore.

[0088] Results and Validation: Samples were taken from the two anomalies for analysis. Anomaly A had a copper grade of 1.2% and anomaly B had a tungsten grade of 0.8%, validating the model's ability to distinguish symbiotic minerals.

[0089] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for identifying deep-earth ore bodies based on deep learning and phased array radar wave velocity inversion, characterized in that, Includes the following steps: S1. The phased array deep-earth detection radar system with a preset frequency is used to perform two-dimensional or three-dimensional scanning of underground space and collect data. S2. Preprocess the data collected in step S1, extract the wave velocity of electromagnetic waves based on tomography, and extract multi-dimensional signal features, including carrier frequency amplitude, signal envelope amplitude, signal attenuation coefficient, and signal spectrum centroid and bandwidth. S3. The wave velocity extracted in step S2 is inverted to obtain the three-dimensional distribution of the dielectric constant, and the three-dimensional distribution of the equivalent conductivity is inverted based on the signal attenuation characteristics. S4. The three-dimensional distribution of dielectric constant, the three-dimensional distribution of equivalent conductivity, and multi-dimensional signal features are used as input features and input into a pre-trained deep neural network model. Then, the model outputs the ore body type label and prediction confidence. S5. Integrate the spatial location, depth, and ore body type identification results to generate a three-dimensional ore body distribution model and output a comprehensive exploration report, which includes ore body type, spatial morphology, burial depth, and confidence level. Step S2 specifically includes: Suppose the underground medium is discretized into N grid cells, and M radar wave ray paths pass through this region, then the... Propagation time of a ray for: ;in, Indicates the first The path length of the ray traveling through the j-th grid. This represents the slowness of the j-th grid. Indicates the first Random errors in the travel time observation of a ray; Write all the ray equations in matrix form, i.e. , where L is A path matrix of dimension s, where s is A slowness vector of dimension t, where t is The observation travel time vector of dimension; The solution is obtained using regularized least squares, i.e. ,in Represents the L2 norm. Represents the regularization parameter. The smoothness constraint matrix is ​​used to obtain the three-dimensional distribution of the slowness vector s, which is then converted into the three-dimensional distribution of the wave velocity v. In step S3, the dielectric constant is... Three-dimensional distribution and equivalent conductivity The three-dimensional distribution is used for full waveform inversion, specifically including: defining the objective function. , Where m represents the model parameter vector. Indicates the first Radar signals measured at each receiving point This represents the result of forward simulation calculations using Maxwell's equations, given m. Radar signals from each receiving point; The full waveform inversion updates the model parameters through iterative optimization, i.e. ,in, Indicates the number of iterations. Indicates step size, Describe the objective function exist gradient at, This represents the inverse of the Hessian matrix.

2. The deep-earth ore body identification method based on deep learning and phased array radar wave velocity inversion according to claim 1, characterized in that: In step S3, using the formula The dielectric constant is obtained by inverting the wave velocity v extracted in step S2. The three-dimensional distribution, and the equivalent conductivity inverted based on the signal attenuation characteristics. The three-dimensional distribution.

3. The deep-earth ore body identification method based on deep learning and phased array radar wave velocity inversion according to claim 2, characterized in that: The architecture of the deep neural network model described in step S4 includes a convolutional neural network (CNN) and a long short-term memory network (E-LSTM).

4. The method for identifying deep-earth ore bodies based on deep learning and phased array radar wave velocity inversion according to claim 3, characterized in that, Step S4 includes the following steps: S41. Construct a feature vector from the features of each spatial location. ,in The characteristic dimensions include the three-dimensional distribution of the dielectric constant at that location, the three-dimensional distribution of the equivalent conductivity, and multi-dimensional signal characteristics. S42. Treat the three-dimensional spatial data volume as a multi-channel image and input it into a deep neural network model. First, perform convolution through the convolutional layer of a convolutional neural network (CNN) to obtain the... Feature map of layer : ,in, Indicates the sequence number of the convolutional layer. This represents the convolution operation. Indicates the first Feature maps input to the layer, Indicates the first The convolutional kernel weights of the layer, Indicates the first Layer bias, This represents the activation function; subsequently, through multiple convolutional and pooling operations of the CNN, spatial features from local to global are extracted. ; S43. Input the spatial features extracted by the CNN into the Long Short-Term Memory (E-LSTM) network, where the E-LSTM network... The internal state at time t is updated as follows: in, This represents the activation value of the forget gate. The weight matrix represents the forget gate. express The input feature vector at time t, This represents the hidden state at time t-1. and Input feature vector at time step To splice, This represents the bias vector of the forget gate. It is the sigmoid activation function; This represents the activation value of the input gate. This represents the weight matrix of the input gate. This represents the bias vector of the input gate; express Cellular state at any given moment express The state of candidate cells at any given time. Indicates calculation The weight matrix, Indicates calculation The bias vector, express t Cell state at time -1; This indicates the activation value of the output gate. express Output at any moment This represents element-wise multiplication. This represents the weight matrix of the output gate. The output gate's bias vector is represented; the gating mechanism based on E-LSTM outputs context features. ; S44. The deep neural network model also includes an online clustering enhancement module, which uses a clustering loss function... Enhance the model for the samples The clustering loss function for: ,in Representing contextual features The features after projection network, where B represents the batch size. Indicates the assignment to the sample The cluster center vector, Indicates sample pseudo-tags; S45, Finally, context features The input is fed into a fully connected layer to calculate the final probability distribution of ore body types: Where K represents the total number of ore body types, and c represents the ore body type number whose probability is to be calculated. This represents the weight vector corresponding to c. This represents the bias term corresponding to c, and m represents the ore body category number used for summation. This represents the weight vector corresponding to m. represents the bias term of m, and y represents the predicted ore body category label. The category with the highest probability is taken as the identification result, and its corresponding maximum probability value is taken as the prediction confidence.

5. The deep-earth ore body identification method based on deep learning and phased array radar wave velocity inversion according to claim 4, characterized in that, Before input features of different dimensions are fed into the pre-trained deep neural network model, step S4 also includes fusing them: ,in, Represents the time-domain feature vector. Represents the frequency domain eigenvector. This represents vector concatenation. This indicates the Z-Score normalization operation. and These represent the weight parameters, This represents the fused feature vector.

Images