A transient electromagnetic near-surface target feature recognition system and method

Abstract

The present invention provides a transient electromagnetic near-surface target feature recognition system and method, relating to the technical field of underground target identification. The system arranges a measurement grid, uses a JPMD sensor to collect the response voltage components of each measurement point, calculates the magnetic gradient tensor, and generates a target position vector. After establishing a global coordinate system, the system uses the Gauss-Newton algorithm to determine the positioning result based on the actual detection value and the single-dipole model prediction value. The target aspect ratio is obtained by decomposing the magnetic gradient tensor, and the decay time coefficient is corrected in combination with soil conductivity to determine the target conductivity. Cluster analysis is performed on the targets to improve classification accuracy. Finally, target recognition models are trained for different clusters to identify the types of targets in the clusters. Real-time targets are assigned to clusters and then identified using the corresponding target recognition models. This method integrates data from multiple measurement points, improving positioning accuracy and target recognition efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of underground target recognition, and in particular to a transient electromagnetic near-surface target feature recognition system and method. Background Art

[0002] In the field of near-surface target detection, especially the detection of underground unexploded ordnance (UXO), traditional methods mainly rely on magnetic exploration or electromagnetic induction technology. However, these technologies often have problems such as insufficient positioning accuracy, poor anti-interference ability and complex data processing. The existing transient electromagnetic method is easily affected by environmental noise under complex geological conditions, resulting in a reduced target recognition rate, and it mostly relies on data from a single measurement point, making it difficult to achieve high-precision three-dimensional positioning. In addition, existing systems usually lack dynamic optimization capabilities and cannot effectively adapt to changes in the electromagnetic characteristics of underground targets. After the target is detected, there is a lack of further identification, which limits its reliability and efficiency in practical applications. Therefore, there is an urgent need for a high-precision transient electromagnetic detection system that can combine data from multiple measurement points and dynamically optimize the positioning algorithm to solve the above technical bottlenecks.

[0003] In the prior art, publication number CN105137498B discloses an underground target detection and identification system and method based on feature fusion, which detects and collects echo signals of ground penetrating radar; preprocesses the collected echo signals; enhances the preprocessed echo signals; extracts the time domain features and wavelet packet energy spectrum features of the enhanced echo signals, and performs Welch power spectrum estimation and high-order spectrum analysis on the enhanced echo signals to obtain Welch power spectrum and high-order spectrum respectively; fuses the time domain features, wavelet packet energy spectrum features, Welch power spectrum and high-order spectrum of the echo signals; performs target recognition on the four types of fused signal features through a wavelet neural network; and outputs the result after target recognition.

[0004] The main problems with the above scheme are: it involves multi-stage feature extraction and fusion, and uses wavelet neural networks for classification and recognition, which leads to excessive computational complexity and insufficient real-time performance in large-scale data scenarios, making it difficult to meet the needs of rapid detection; and the simultaneous integration of time domain features, wavelet packet energy spectrum features, Welch power spectrum and high-order spectrum results in information overlap and feature redundancy; ground penetrating radar is sensitive to shallow high-dielectric constant targets, and has limited detection capabilities for deep small targets or low-dielectric constant targets, which limits the scope of application of the scheme.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0006] The object of the present invention is to provide a transient electromagnetic near-surface target feature recognition system and method to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A transient electromagnetic near-surface target feature recognition system, specifically comprising:

[0009] A data acquisition module is used to arrange a 3×3 measurement grid in the target area, use the center point of each grid as a measurement point, and collect the response voltage component of each measurement point;

[0010] an initial position measurement module, configured to calculate a magnetic gradient tensor of each measurement point based on a response voltage component of each measurement point, calculate a target position vector of each measurement point based on the magnetic gradient tensor and the response voltage component, establish a global coordinate system in the measurement grid, generate a global target position of the measurement point based on the target position vectors of all measurement points, and average the global target positions of all measurement points to generate an average global target position;

[0011] A model building module is used to construct a single dipole model at each measurement point, equating the underground target to a magnetic dipole with a dipole moment, and predicting the response voltage component through the single dipole model;

[0012] The initial position optimization and positioning module is used to use the average global target position as the initial value of the Gauss-Newton algorithm, construct a residual matrix reflecting the difference between the actual response voltage and the response voltage predicted by the single dipole model, calculate the negative derivative of the residual matrix to generate the Jacobian matrix, iterate from the initial value to update the average global target position until the maximum number of iterations is reached, and output the last updated average global target position as the positioning result;

[0013] The target feature extraction module is used to obtain and decompose the magnetic gradient tensor in the area where the positioning result is located, generate the target aspect ratio, and simultaneously measure the soil conductivity. The decay time coefficient of the transient electromagnetic response is corrected by the soil conductivity to generate the actual decay time coefficient, which reflects the conductivity of the target.

[0014] The target clustering module is used to collect the aspect ratios and actual decay time coefficients of known targets appearing near the surface in historical data, perform Min-Max normalization on the aspect ratios and actual decay time coefficients, generate the feature vector of each known target, calculate the Euclidean distance between all known target feature vectors, take the 30% quantile of the Euclidean distance as the neighborhood radius, and set the minimum number of samples for clustering. Unsupervised cluster analysis is performed on the feature vector of each known target based on the DBSCAN clustering algorithm. The feature vector of each known target in the cluster is equivalent to a sample point. The mean of the sample points in each cluster is calculated as the cluster center, and the Mahalanobis distance from each sample point in the cluster to the cluster center is generated. A Mahalanobis distance threshold is set. When the Mahalanobis distance from a sample point to the cluster center exceeds the Mahalanobis distance threshold, the sample point is removed from the cluster. After each group of sample points is removed, the cluster center is recalculated, and the Mahalanobis distances of other sample points to the cluster center are calculated. The above steps of removing sample points are repeated until there are no sample points in the cluster whose Mahalanobis distance from the cluster center exceeds the Mahalanobis distance threshold.

[0015] The target recognition module is used to build a deep learning network. It takes the known target feature vectors in different clusters as input and the category of the known target as the label. It trains the target recognition model for each cluster, calculates the feature vector of the real-time located target according to the above steps, selects the cluster center with the closest Euclidean distance to the real-time located target feature vector, assigns the real-time located target feature vector to the cluster corresponding to the cluster center, and inputs the real-time located target feature vector into the target recognition model of the corresponding cluster to identify the category of the real-time located target.

[0016] Furthermore, a JPMD sensor consisting of a single-layer transmitting coil and five three-axis receiving coils was used for data acquisition. The spacing of the measurement grid was 25 cm, and 15 response voltage components were collected at each measurement point.

[0017] Furthermore, in the JPMD sensor, the center of the JPMD sensor is used as the origin of the coordinate system. There is a three-axis receiving coil at the origin of the coordinate system, and the other four three-axis receiving coils are distributed in the four positive directions of front, back, left, and right of the origin of the coordinate system. The positive direction of the x-axis is the direction pointing to the right of the origin of the coordinate system, the positive direction of the y-axis is the direction pointing to the rear of the origin of the coordinate system, and the positive direction of the z-axis is perpendicular to the x-axis and y-axis and passes through the origin of the coordinate system and points upward. A measurement point coordinate system is established, and the three-axis receiving coil at the origin of the coordinate system is set to R3, the coils in the positive and negative directions of the x-axis are set to R2 and R4 respectively, and the coils in the positive and negative directions of the y-axis are set to R1 and R5 respectively.

[0018] Furthermore, the principle for calculating the target position vector of each measurement point is as follows:

[0019] The magnetic gradient tensor at the measurement point is:

[0020]

[0021] Where G represents the magnetic gradient tensor of the i-th measurement point, i represents the index of the measurement point, and i∈[1,9], g xx,i Indicates the rate of change of the magnetic field in the x direction at the i-th measurement point, g xy,i Indicates the rate of change of the magnetic field in the y direction at the i-th measurement point in the x direction, g xz,i Indicates the rate of change of the magnetic field in the z direction in the x direction at the i-th measurement point, g yx,i Indicates the rate of change of the magnetic field in the x-direction to the y-direction at the i-th measurement point, g yy,i Indicates the rate of change of the magnetic field in the y direction at the i-th measurement point, g yz,i Indicates the rate of change of the magnetic field in the z direction in the y direction at the i-th measurement point, g zx,i Indicates the rate of change of the magnetic field in the x-direction in the z-direction at the i-th measurement point, g zy,i Indicates the rate of change of the magnetic field in the y direction in the z direction at the i-th measurement point, g zz,i represents the rate of change of the magnetic field in the z direction at the i-th measurement point;

[0022]

[0023] g zz,i =-(g xx,i +g yy,i )

[0024] Among them, V 4x,i represents the x-axis response voltage component of R4 at the i-th measurement point, V 2x,i represents the x-axis response voltage component of R2 at the i-th measurement point, d represents the distance from R1 to the origin of the coordinate system, V 4y,i represents the y-axis response voltage component of R4 at the i-th measurement point, V 2y,i Represents the y-axis response voltage component of R2 at the i-th measurement point, V 4z,i represents the z-axis response voltage component of R4 at the i-th measurement point, V 2z,i represents the z-axis response voltage component of R2 at the i-th measurement point, V 1y,i Represents the y-axis response voltage component of R1 at the i-th measurement point, V 5y,i Represents the y-axis response voltage component of R5 at the i-th measurement point, V 1z,i represents the z-axis response voltage component of R1 at the i-th measurement point, V 5z,i represents the z-axis response voltage component of R5 in the i-th measurement point;

[0025] The formula for calculating the target position vector of each measurement point is:

[0026]

[0027] Among them, R i Represents the target position vector of the i-th measurement point, V i Represents the response voltage component of the i-th measurement point.

[0028] Furthermore, the principle for generating the average global target position is:

[0029] The center point of the 3×3 measurement grid is used as the global origin. On the plane where the measurement grid is located, there is a measurement point in each of the eight positive directions of the global origin: up, down, left, right, upper left, lower left, upper right, and lower right. The global x-axis is defined as the positive direction of the global origin pointing due right, the global y-axis is defined as the positive direction of the global origin pointing due upward, and the global z-axis is defined as the positive direction passing through the global origin, perpendicular to the global x-axis and global y-axis, and pointing outside the grid plane. The global coordinate system of the target area is constructed.

[0030] For each measurement point, there is a known global coordinate and target position vector, then the global target position of a single measurement point is:

[0031] r i =S i +R i

[0032] Among them, r i represents the global target position of the i-th measurement point, S i represents the global coordinates of the i-th measurement point;

[0033] The average global target position is:

[0034]

[0035] Among them, r d represents the average global target position.

[0036] Furthermore, the principle for updating the target position is:

[0037] The residual matrix is:

[0038] F(r d )=V obs -V pre

[0039] Among them, F(r d ) represents the residual matrix, V obs Represents the response voltage at the measurement point, V pre represents the response voltage predicted by the single-dipole model;

[0040] The Jacobian matrix is:

[0041]

[0042] Where J represents the Jacobian matrix;

[0043] The iterative equation for the target position is:

[0044] r d k+1 =r d k +Δr d k

[0045] Δr d k =(J T J) -1 J T F(r d k )

[0046] Among them, r d k+1 represents the average global target position at the k+1th iteration, r d k represents the average global target position of the kth iteration, k represents the index of the number of iterations, and k∈[0,K], K represents the maximum number of iterations. When k=0, r d 0 =r d , r d represents the initial average global target position before iteration, Δr d k Represents the correction amount to the average global target position at the kth iteration.

[0047] Furthermore, the principle for generating the target aspect ratio is as follows:

[0048] The magnetic gradient tensor in the area where the positioning result is located is:

[0049]

[0050] Among them, G0 represents the magnetic gradient tensor of the area where the positioning result is located, B x 、B y 、B z Represent the magnetic field components of the x, y, and z axes respectively;

[0051] The formula for decomposing G0 to obtain three eigenvalues ​​is:

[0052] Solve the equation |G0-λ i I|=0, we get three eigenvalues ​​λ1, λ2, and λ3, and |λ1|>|λ2|>|λ3|, where I represents the identity matrix;

[0053] The aspect ratio is:

[0054]

[0055] Where L represents the target aspect ratio.

[0056] Furthermore, the principle for judging the conductivity of the target based on the corrected decay time coefficient is as follows:

[0057] The formula for correcting the decay time coefficient is:

[0058] κ=κ0(1+kσ s )

[0059] Among them, κ represents the actual decay time coefficient, κ0 represents the decay time coefficient before correction, k represents the empirical coefficient, σ s Indicates soil electrical conductivity;

[0060] κ is proportional to the target conductivity.

[0061] Furthermore, the principle of unsupervised cluster analysis is:

[0062] The formula for calculating the Euclidean distance between all known target feature vectors is:

[0063]

[0064] Among them, v o 、v p denote the oth and pth known target feature vectors respectively, d(v o ,v p ) means v o 、v p The Euclidean distance between them, o, p∈[1,P] and o≠p, P represents the total number of known target feature vectors, L′ o represents the normalized aspect ratio of the oth known target feature vector, L′ p represents the normalized aspect ratio of the pth known target feature vector, κ′ o represents the normalized actual decay time coefficient of the oth known target feature vector, κ′ p represents the normalized actual decay time coefficient of the p-th known target feature vector;

[0065] The neighborhood radius of the cluster is taken as the 30% quantile of the Euclidean distance of all sample points, and the minimum number of sample points for the cluster is Indicates taking the square root of the total number of known target feature vectors and then rounding up; randomly select a sample point and count the other sample points within the neighborhood radius. If the number exceeds The selected sample points are regarded as cluster centers, and all sample points are assigned to the cluster with the closest Euclidean distance to the cluster center, which is considered to be the completion of unsupervised cluster analysis;

[0066] The cluster center is calculated in each cluster according to the formula:

[0067]

[0068] Among them, c q represents the cluster center of the qth cluster, q represents the cluster index, P q represents the number of sample points in the qth cluster, j represents the index of the sample point in the cluster, and j∈[1,J q ], J q represents the number of sample points in the qth cluster, v j Represents the target feature vector corresponding to the j-th sample point;

[0069] The formula for generating the Mahalanobis distance from each eigenvector in a cluster to the cluster center is:

[0070]

[0071] Among them, D j,q S represents the Mahalanobis distance from the jth sample point to the cluster center in the qth cluster. q represents the intra-class covariance matrix of the qth cluster;

[0072] When D j,q >D max When , the sample point is removed from the cluster, D max Represents the Mahalanobis distance threshold.

[0073] The present invention also provides a transient electromagnetic near-surface target feature recognition method, which is performed by the above-mentioned transient electromagnetic near-surface target feature recognition system, and specifically comprises the following steps:

[0074] Step 1: Arrange a 3×3 measurement grid in the target area, use the center point of each grid as a measurement point, and collect the response voltage component of each measurement point;

[0075] Step 2: Calculate the magnetic gradient tensor of each measurement point based on the response voltage component of each measurement point, calculate the target position vector of each measurement point based on the magnetic gradient tensor and the response voltage component, establish a global coordinate system in the measurement grid, generate the global target position of the measurement point based on the target position vectors of all measurement points, and average the global target positions of all measurement points to generate an average global target position;

[0076] Step 3: Construct a single dipole model at each measurement point, equating the underground target to a magnetic dipole with a dipole moment, and predict the response voltage using the single dipole model;

[0077] Step 4: Use the average global target position as the initial value of the Gauss-Newton algorithm, construct a residual matrix that reflects the difference between the actual response voltage and the response voltage predicted by the single dipole model, take the negative derivative of the residual matrix to generate the Jacobian matrix, and iterate from the initial value to update the average global target position until the maximum number of iterations is reached. Output the last updated average global target position as the positioning result;

[0078] Step 5: Obtain and decompose the magnetic gradient tensor in the area where the positioning result is located, generate the target aspect ratio, and simultaneously measure the soil conductivity. Use the soil conductivity to correct the decay time coefficient of the transient electromagnetic response to generate the actual decay time coefficient, which reflects the conductivity of the target.

[0079] Step 6: Collect the aspect ratios and actual decay time coefficients of known targets appearing near the surface in historical data, perform Min-Max normalization on the aspect ratios and actual decay time coefficients, generate the feature vector of each known target, calculate the Euclidean distance between all feature vectors of known targets, take the 30% quantile of the Euclidean distance as the neighborhood radius, and set the minimum number of samples for clustering. Perform unsupervised cluster analysis on the feature vector of each known target based on the DBSCAN clustering algorithm. The feature vector of each known target in the cluster is equivalent to a sample point. In each cluster, the mean of the sample points is calculated as the cluster center. The Mahalanobis distance from each sample point in the cluster to the cluster center is generated, and a Mahalanobis distance threshold is set. When the Mahalanobis distance from a sample point to the cluster center exceeds the Mahalanobis distance threshold, the sample point is removed from the cluster. After each group of sample points is removed, the cluster center is recalculated, and the Mahalanobis distances of other sample points to the cluster center are calculated. Repeat the above steps of removing sample points until there are no sample points in the cluster whose Mahalanobis distance from the cluster center exceeds the Mahalanobis distance threshold.

[0080] Step 7: Build a deep learning network, take the known target feature vectors in different clusters as input, and the category of the known target as the label, train the target recognition model for each cluster, calculate the feature vector of the real-time located target according to the above steps, select the cluster center with the closest Euclidean distance to the real-time located target feature vector, assign the real-time located target feature vector to the cluster corresponding to the cluster center, and input the real-time located target feature vector into the target recognition model of the corresponding cluster to identify the category of the real-time located target.

[0081] Compared with the prior art, the present invention has the following beneficial effects:

[0082] By calculating the magnetic gradient tensor at each measurement point, the present invention can more accurately derive the target position vector. Compared to traditional single-point response data, this magnetic field gradient-based positioning method can effectively reduce the impact of environmental noise and measurement errors, improve target positioning accuracy, and achieve effective fusion of data from each measurement point. This method can produce more consistent global positioning results, help eliminate local measurement errors, and improve the credibility of the overall data.

[0083] The present invention also constructs a single-dipole model for the measurement point, which can effectively simulate the electromagnetic response of small near-surface targets and has a wide range of applications. Compared with complex multi-dipole models, the single-dipole model can still provide relatively accurate predictions under conditions of small amounts of data, reducing computational complexity. By constructing a residual matrix that reflects the difference between actual data and model prediction results, the shortcomings of the model are clearly identified, and real-time, continuous optimization is achieved. While ensuring accuracy, computational complexity is reduced and operational efficiency is improved. The present invention also obtains and decomposes the magnetic gradient tensor, which can not only generate the aspect ratio of the target, but also extract the geometric features of the target, providing more comprehensive target information. By correcting the decay time coefficient of the transient electromagnetic response through soil conductivity, the true electrical characteristics of the target can be accurately reflected, reducing the impact of environmental factors on target identification and improving the accuracy of judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 This is a schematic diagram of the system modules of an embodiment of the present invention;

[0085] Figure 2 This is a line graph of the soil electrical conductivity correction decay time coefficient according to an embodiment of the present invention;

[0086] Figure 3 Schematic diagram of the method flow of an embodiment of the present invention. DETAILED DESCRIPTION

[0087] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to specific embodiments.

[0088] It should be noted that, unless otherwise defined, the technical or scientific terms used in the present invention should have the usual meanings understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative position relationships. When the absolute position of the object being described changes, the relative position relationship may also change accordingly.

[0089] Example:

[0090] See also Figures 1 to 2 , the present invention provides a technical solution:

[0091] A transient electromagnetic near-surface target feature recognition system, specifically comprising:

[0092] A data acquisition module is used to arrange a 3×3 measurement grid in the target area, use the center point of each grid as a measurement point, and collect the response voltage component of each measurement point;

[0093] In this embodiment, a JPMD sensor including a single-layer transmitting coil and five three-axis receiving coils is used for data acquisition. The spacing of the measurement grid is 25 cm, and 15 response voltage components are collected at each measurement point.

[0094] The transmitting coil of the JPMD sensor has a diameter of 50 cm, a height of 7 cm, and 30 turns. The height of the five receiving coils is 6 cm, and the distance between the receiving coil located at the center of the transmitting coil and the other receiving coils is 20 cm.

[0095] In the JPMD sensor, the center of the JPMD sensor is used as the origin of the coordinate system. There is a three-axis receiving coil at the origin of the coordinate system, and the other four three-axis receiving coils are distributed in the four positive directions of the coordinate system origin, namely, front, back, left, and right. The positive direction of the x-axis is the direction pointing to the right of the coordinate system origin, the positive direction of the y-axis is the direction pointing to the back of the coordinate system origin, and the positive direction of the z-axis is perpendicular to the x-axis and y-axis, passes through the origin of the coordinate system, and points upward. A measurement point coordinate system is established, and the three-axis receiving coil at the origin of the coordinate system is set as R3, the coils in the positive and negative directions of the x-axis are set as R2 and R4, and the coils in the positive and negative directions of the y-axis are set as R1 and R5, respectively.

[0096] an initial position measurement module, configured to calculate a magnetic gradient tensor of each measurement point based on a response voltage component of each measurement point, calculate a target position vector of each measurement point based on the magnetic gradient tensor and the response voltage component, establish a global coordinate system in the measurement grid, generate a global target position of the measurement point based on the target position vectors of all measurement points, and average the global target positions of all measurement points to generate an average global target position;

[0097] In this embodiment, the principle for calculating the target position vector of each measurement point is:

[0098] The magnetic gradient tensor at the measurement point is:

[0099]

[0100] Where G represents the magnetic gradient tensor of the i-th measurement point, i represents the index of the measurement point, and i∈[1,9], g xx,i Indicates the rate of change of the magnetic field in the x direction at the i-th measurement point, g xy,i Indicates the rate of change of the magnetic field in the y direction at the i-th measurement point in the x direction, g xz,i Indicates the rate of change of the magnetic field in the z direction in the x direction at the i-th measurement point, g yx,i Indicates the rate of change of the magnetic field in the x-direction to the y-direction at the i-th measurement point, g yy,i Indicates the rate of change of the magnetic field in the y direction at the i-th measurement point, g yz,i Indicates the rate of change of the magnetic field in the z direction in the y direction at the i-th measurement point, g zx,i Indicates the rate of change of the magnetic field in the x-direction in the z-direction at the i-th measurement point, g zy,i Indicates the rate of change of the magnetic field in the y direction in the z direction at the i-th measurement point, g zz,i represents the rate of change of the magnetic field in the z direction at the i-th measurement point;

[0101] The magnetic gradient tensor is a 3×3 matrix that reflects the rate of change of the magnetic field in the x, y, and z directions. The calculation of the magnetic gradient tensor is based on the response voltage measured by multiple receiving coils. Each receiving coil measures the response voltage in the x, y, and z directions. The rate of change of the magnetic field in each direction is calculated using the differential approximation method.

[0102]

[0103] g zz,i =-(g xx,i +g yy,i )

[0104] Among them, V 4x,i represents the x-axis response voltage component of R4 at the i-th measurement point, V 2x,irepresents the x-axis response voltage component of R2 at the i-th measurement point, d represents the distance from R1 to the origin of the coordinate system, V 4y,i represents the y-axis response voltage component of R4 at the i-th measurement point, V 2y,i Represents the y-axis response voltage component of R2 at the i-th measurement point, V 4z,i represents the z-axis response voltage component of R4 at the i-th measurement point, V 2z,i represents the z-axis response voltage component of R2 at the i-th measurement point, V 1y,i Represents the y-axis response voltage component of R1 at the i-th measurement point, V 5y,i Represents the y-axis response voltage component of R5 at the i-th measurement point, V 1z,i represents the z-axis response voltage component of R1 at the i-th measurement point, V 5z,i represents the z-axis response voltage component of R5 in the i-th measurement point;

[0105] In a uniform grid, the rate of change of the magnetic field, g, is calculated based on the first-order central difference formula. xx,i It represents the rate of change of the magnetic field in the x-direction at the measurement point. To calculate this rate of change, two symmetrical points need to be selected on the x-axis. The difference between the response voltage component measurements between the symmetrical points and the distance between them can be approximated to obtain the rate of change of the magnetic field. The distances of R1, R2, R4, and R5 from the origin of the coordinate system are all the same, d. R4 is located in the positive direction of the x-axis, and the distance from the origin is d. R2 is located in the negative direction of the x-axis, and the distance from the origin is d. Therefore, the magnetic field components V of R4 and R2 in the x-axis direction are 4x,i and V 2x,i It directly reflects the change of the magnetic field in the x-axis direction, so the calculation of g xx,i The difference formula is The magnetic field is a non-divergent field, and its divergence is zero, so g xx,i +g yy,i +g zz,i =0,g zz,i =-(g xx,i +g yy,i ), g zz,i No independent measurement is required, it can be directly measured by g xx,i and g yy,i Calculation shows that in low-frequency electromagnetic fields, the magnetic gradient tensor is symmetrical, that is, g xy,i =g yx,i , g xz,i =g zx,i , g yz,i =g zy,i ; Calculate g xy,i When using g directly yx,i The calculation formula is the rate of change of the magnetic field in the x direction in the y direction. Similarly, g is calculated xz,i 、g yy,i 、gyz,i .

[0106] The formula for calculating the target position vector of each measurement point is:

[0107]

[0108] Among them, R i Represents the target position vector of the i-th measurement point, V i Represents the response voltage component of the i-th measurement point.

[0109] The formula for calculating the target position vector is based on the magnetic dipole model. The magnetic field generated by the target is approximately a dipole field. Through the linear relationship between the magnetic gradient tensor and the response voltage, the position vector of the target relative to the measurement point is inverted, that is, R i , the underground target is approximated as a point magnetic dipole, and the magnetic field generated by it is distributed in space to meet Where B(r) represents the magnetic field generated by the target, m represents the magnetic dipole moment of the target, and r represents the position vector from the target to the sensor. represents the unit direction vector, μ0 represents the vacuum permeability, in the far field, that is, the target distance is much larger than the dipole size, then In the transient electromagnetic method, the response voltage of the receiving coil is proportional to the rate of change of the magnetic field, V i ∝G i m, the magnetic moment m is proportional to the external excitation magnetic field and the polarizability of the target. The external excitation magnetic field is the magnetic field generated by the transmitting coil, V i ∝G i ·r, take the inverse of both sides, then The proportionality factor is approximately -3, then The above formula is based on the observed electromagnetic response V i Inversely calculate the target position.

[0110] The principle behind generating the average global target position is:

[0111] The center point of the 3×3 measurement grid is used as the global origin. On the plane where the measurement grid is located, there is a measurement point in each of the eight positive directions of the global origin: up, down, left, right, upper left, lower left, upper right, and lower right. The global x-axis is defined as the positive direction of the global origin pointing due right, the global y-axis is defined as the positive direction of the global origin pointing due upward, and the global z-axis is defined as the positive direction passing through the global origin, perpendicular to the global x-axis and global y-axis, and pointing outside the grid plane. The global coordinate system of the target area is constructed.

[0112] For each measurement point, there is a known global coordinate and target position vector, then the global target position of a single measurement point is:

[0113] r i=S i +R i

[0114] Among them, r i represents the global target position of the i-th measurement point, S i represents the global coordinates of the i-th measurement point;

[0115] Target position vector R i Indicates the offset of the target relative to the current measurement point sensor position, S i It is the coordinate of the measurement point in the global coordinate system, which reflects the position of the sensor in the global space. The position of the target in the global coordinate system = the global position of the sensor + the offset of the target relative to the sensor. The local coordinate system of the sensor is kept completely consistent with the axial direction of the global coordinate system. Therefore, S i and R i Direct addition, R i The essence of is the vector pointing from the sensor position to the target. The essence of adding the two is to offset the target in the local coordinate system R i Converted to the global coordinate system, in the global coordinate system, the absolute position of the target is the sensor position plus this vector, by r i Taking an average can suppress single-point measurement errors and improve positioning robustness.

[0116] The average global target position is:

[0117]

[0118] Among them, r d represents the average global target position.

[0119] A model building module is used to construct a single dipole model at each measurement point, equating the underground target to a magnetic dipole with a dipole moment, and predicting the response voltage component through the single dipole model;

[0120] In this embodiment, the technical features of predicting the response voltage using a single dipole model have been disclosed in the paper “Fast Localization of Underground Targets by Magnetic Gradient Tensor and Gaussian-Newton Algorithm With a Portable Transient Electromagnetic System LiJie Wangetc.” and are not described in detail here.

[0121] The initial position optimization and positioning module is used to use the average global target position as the initial value of the Gauss-Newton algorithm, construct a residual matrix reflecting the difference between the actual response voltage and the response voltage predicted by the single dipole model, take the negative derivative of the residual matrix to generate the Jacobian matrix, iterate from the initial value to update the target position until the maximum number of iterations is reached, and output the last updated target position as the positioning result;

[0122] In this embodiment, the principle for updating the target position is:

[0123] The residual matrix is:

[0124] F(r d )=V obs -V pre

[0125] Among them, F(r d ) represents the residual matrix, V obs Represents the response voltage at the measurement point, V pre represents the response voltage predicted by the single-dipole model;

[0126] The core of the residual matrix is ​​to quantify the difference between the actual observation data and the model prediction data. During the target positioning process, there is a deviation between the response voltage predicted by the single dipole model and the response voltage measured by the actual sensor. The residual matrix gradually adjusts the estimated value of the target position by calculating this deviation to change the response voltage measured by the actual sensor until the estimated value of the target position is close to the actual position, V obs Represents the response voltage data actually measured by the sensor, which is a multidimensional vector containing multiple measurement points and coil components. pre The theoretical voltage calculated by the single dipole model is smaller. The smaller the residual, the closer the actual voltage is to the theoretical voltage. The average global target position calculated is closer to the actual positioning result. The voltage V predicted by the single dipole model is pre Depends on the average global target position r d , when r d When changes, the predicted magnetic field distribution and response voltage change, resulting in a change in the residual;

[0127] V obs Represents the response voltage, which is a matrix composed of the response voltage components of the measurement point. V pre is the matrix composed of the response voltage components predicted by the single dipole model; the matrix composed of the response voltage components of the i-th measurement point is composed of the response voltage components of the five receiving coils in the x, y, and z directions. Among them, V i Represents the response voltage component matrix of the i-th measurement point, V 1x,i To V 5x,iRespectively represent the response voltage components of R1 to R5 in the x-axis direction, V 1y,i To V 5y,i Respectively represent the response voltage components of R1 to R5 in the y-axis direction, V 1z,i To V 5z,i Represent the response voltage components of R1 to R5 in the z-axis direction respectively, and the matrix of all measurement points is combined to generate V obs , V pre It is also in this form.

[0128] The Jacobian matrix is:

[0129]

[0130] Where J represents the Jacobian matrix;

[0131] The Jacobian matrix is ​​used to describe the sensitivity of the residual to the target position, that is, how the residual changes with the target position. The Jacobian matrix is ​​the negative partial derivative matrix of the residual matrix with respect to the target position. The negative sign indicates that the position is updated in the direction of decreasing residual. Each row in the Jacobian matrix corresponds to the residual of a measurement point, and each column corresponds to the partial derivative of the target position.

[0132] The iterative equation for the target position is:

[0133] r d k+1 =r d k +Δr d k

[0134] Δr d k =(J T J) -1 J T F(r d k )

[0135] Among them, r d k+1 represents the average global target position at the k+1th iteration, r d k represents the average global target position of the kth iteration, k represents the index of the number of iterations, and k∈[0,K], K represents the maximum number of iterations. When k=0, r d 0 =r d , r d represents the initial average global target position before iteration, Δr d k Represents the correction amount to the average global target position at the kth iteration.

[0136] The target feature extraction module is used to obtain and decompose the magnetic gradient tensor in the area where the positioning result is located, generate the target aspect ratio, and simultaneously measure the soil conductivity. The decay time coefficient of the transient electromagnetic response is corrected by the soil conductivity to generate the actual decay time coefficient, which reflects the conductivity of the target.

[0137] In this embodiment, the principle for generating the target aspect ratio is:

[0138] The magnetic gradient tensor in the area where the positioning result is located is:

[0139]

[0140] Among them, G0 represents the magnetic gradient tensor of the area where the positioning result is located, B x 、B y 、B z Represent the magnetic field components of the x, y, and z axes respectively;

[0141] The formula for decomposing G0 to obtain three eigenvalues ​​is:

[0142] Based on the eigenvalue decomposition of the matrix, solve the equation |G0-λ i I|=0, we get three eigenvalues ​​λ1, λ2, and λ3, and |λ1|>|λ2|>|λ3|, where I represents the identity matrix;

[0143] The aspect ratio is:

[0144]

[0145] Where L represents the target aspect ratio.

[0146] The eigenvalue decomposition of the magnetic gradient tensor is essentially to decouple the distribution of the magnetic field gradient in space into three orthogonal principal axes. Each eigenvalue corresponds to a direction in which the magnetic field gradient changes most significantly. The three eigenvalues ​​reflect the geometric and physical properties of the target magnetic source. The maximum eigenvalue λ1 represents the rate of change of the magnetic field gradient in the direction of the longest principal axis of the target, which is positively correlated with the length or extension direction of the target. The intermediate eigenvalue λ2 reflects the rate of change of the magnetic field gradient in the direction of the secondary principal axis of the target, which is used to reflect the symmetry of the target. The minimum eigenvalue λ3 represents the rate of change of the magnetic field gradient in the direction of the shortest principal axis of the target. The shape of the target can be roughly reflected by λ1 and λ3. If the target is long and narrow, λ1 is significantly larger than the other two eigenvalues. If the target is flat, λ3 will be significantly smaller than the other two eigenvalues. To reflect the aspect ratio of the target, the anisotropy of the target is quantified. When L>>1, the target is a long strip, which meets the shape characteristics of objects such as cables and unexploded ordnance. When L≈1, the target is close to a sphere or cube, which meets the shape characteristics of objects such as mines and metal blocks.

[0147] The principle of judging the conductivity of a target based on the corrected decay time coefficient is as follows:

[0148] The formula for correcting the decay time coefficient is:

[0149] k=κ0(1+kσ s )

[0150] Among them, κ represents the actual decay time coefficient, κ0 represents the decay time coefficient before correction, k represents the empirical coefficient, σ s Indicates soil electrical conductivity;

[0151] κ is proportional to the target conductivity.

[0152] In transient electromagnetic response, the decay time coefficient of the signal depends not only on the conductivity of the target itself, but also on the conductivity of the surrounding soil. Highly conductive soil will absorb part of the electromagnetic energy due to the eddy current effect, resulting in the observed decay time coefficient being smaller than the true value. Therefore, it is necessary to dynamically correct the soil conductivity. In most surface environments, the effect of soil conductivity on the decay time coefficient is linear. Through the linear correction model (1+kσ s ) is used to correct κ0, which reflects the positive influence of soil conductivity on the correction of the decay time coefficient, that is, the higher the soil conductivity, the greater the correction of the decay time coefficient to balance the influence of soil conductivity; the empirical coefficient k quantifies the correction strength of soil conductivity on the decay time coefficient, and the decay time coefficient of the target is measured under different soil conductivity conditions, and the optimal value of k is determined by regression analysis; the decay time coefficient represents the time required for the transient electromagnetic signal to decay from an initial value to a certain proportion, reflecting the speed of eddy current dissipation capacity. The larger the decay time coefficient, the longer the decay time and the eddy current. The slower the eddy current decays, the better the target conductivity. Conversely, the shorter the decay time, the faster the eddy current decays, and the worse the target conductivity. Targets with high conductivity, such as copper and aluminum, have slow eddy current decay and a larger κ. Low conductivity targets, such as rocks, have slow eddy current decay and a smaller κ. Large or high-aspect-ratio targets may exhibit longer decay times due to eddy current path delays. The corrected κ eliminates the influence of soil conductivity, more directly reflects the target's true conductivity, improves target recognition accuracy, and avoids misjudgments due to environmental changes. The soil conductivity correction is shown in Table 1.

[0153] Table 1.

[0154]

[0155]

[0156] The target clustering module is used to collect the aspect ratios and actual decay time coefficients of known targets appearing near the surface in historical data, perform Min-Max normalization on the aspect ratios and actual decay time coefficients, generate the feature vector of each known target, calculate the Euclidean distance between all known target feature vectors, take the 30% quantile of the Euclidean distance as the neighborhood radius, and set the minimum number of samples for clustering. Unsupervised cluster analysis is performed on the feature vector of each known target based on the DBSCAN clustering algorithm. The feature vector of each known target in the cluster is equivalent to a sample point. The mean of the sample points in each cluster is calculated as the cluster center, and the Mahalanobis distance from each sample point in the cluster to the cluster center is generated. A Mahalanobis distance threshold is set. When the Mahalanobis distance from a sample point to the cluster center exceeds the Mahalanobis distance threshold, the sample point is removed from the cluster. After each group of sample points is removed, the cluster center is recalculated, and the Mahalanobis distances of other sample points to the cluster center are calculated. The above steps of removing sample points are repeated until there are no sample points in the cluster whose Mahalanobis distance from the cluster center exceeds the Mahalanobis distance threshold.

[0157] In this embodiment, the principle of performing unsupervised cluster analysis is:

[0158] The formula for calculating the Euclidean distance between all known target feature vectors is:

[0159]

[0160] Among them, v o 、v p denote the oth and pth known target feature vectors respectively, d(v o ,v p ) means v o 、v p The Euclidean distance between them, o, p∈[1,P] and o≠p, P represents the total number of known target feature vectors, L′ o represents the normalized aspect ratio of the oth known target feature vector, L′ p represents the normalized aspect ratio of the pth known target feature vector, k′ o represents the normalized actual decay time coefficient of the oth known target feature vector, k′ p represents the normalized actual decay time coefficient of the p-th known target feature vector;

[0161] The neighborhood radius of the cluster is taken as the 30% quantile of the Euclidean distance of all sample points, and the minimum number of sample points for the cluster is Indicates taking the square root of the total number of known target feature vectors and then rounding up; randomly select a sample point and count the other sample points within the neighborhood radius. If the number exceeds The selected sample points are regarded as cluster centers, and all sample points are assigned to the cluster with the closest Euclidean distance to the cluster center, which is considered to be the completion of unsupervised cluster analysis;

[0162] The cluster center is calculated in each cluster according to the formula:

[0163]

[0164] Among them, c q represents the cluster center of the qth cluster, q represents the cluster index, P q represents the number of sample points in the qth cluster, j represents the index of the sample point in the cluster, and j∈[1,J q ], J q represents the number of sample points in the qth cluster, v j Represents the target feature vector corresponding to the j-th sample point;

[0165] The formula for generating the Mahalanobis distance from each eigenvector in a cluster to the cluster center is:

[0166]

[0167] Among them, D j,q S represents the Mahalanobis distance from the jth sample point to the cluster center in the qth cluster. q represents the intra-class covariance matrix of the qth cluster;

[0168] When D j,q >D max When , the sample point is removed from the cluster, D max Represents the Mahalanobis distance threshold.

[0169] The unsupervised cluster analysis described above is designed to ensure more accurate grouping of possible near-surface targets based on known target characteristics. Targets with similar electromagnetic characteristics are grouped together to identify different types of near-surface targets. In the absence of prior labels, clustering is used to discover natural distribution patterns in the data, avoiding the subjectivity of human classification. After clustering is generated, a small sample analysis is performed. The Mahalanobis distance is used to evaluate the degree of deviation between sample points and the cluster center within each cluster. Those with excessive deviation are removed, and multiple iterations are performed. After each batch of sample points is removed, the cluster center is recalculated until no sample points in the cluster deviate from the cluster center by more than the Mahalanobis clustering threshold, ensuring that each cluster contains only targets with consistent characteristics. Compared to the Euclidean distance, the Mahalanobis distance takes into account the correlation between features. That is, the aspect ratio of an object may affect conductivity. The aspect ratio reflects the geometric characteristics of the target, and the decay time coefficient reflects conductivity. Combining the two can distinguish targets with similar shapes but different materials. Cluster analysis uses multi-dimensional features to comprehensively judge, avoiding the limitations of a single feature, ensuring that the generated clusters truly reflect that targets with similar feature vectors can be classified into the same category.

[0170] The target recognition module is used to build a deep learning network. It takes the known target feature vectors in different clusters as input and the category of the known target as the label. It trains the target recognition model for each cluster, calculates the feature vector of the real-time located target according to the above steps, selects the cluster center with the closest Euclidean distance to the real-time located target feature vector, assigns the real-time located target feature vector to the cluster corresponding to the cluster center, and inputs the real-time located target feature vector into the target recognition model of the corresponding cluster to identify the category of the real-time located target.

[0171] In this embodiment, the structure of the deep learning network of each cluster is:

[0172] Input layer: contains 1 node, used to input the target feature vector;

[0173] The first hidden layer contains 64 neurons and uses ReLU as the activation function;

[0174] The second hidden layer contains 32 neurons and uses ReLU as the activation function;

[0175] Output layer: contains 1 node, used to output the target category.

[0176] See also Figure 3 The present invention also provides a transient electromagnetic near-surface target feature recognition method, which is executed by the above-mentioned transient electromagnetic near-surface target feature recognition system, and specifically comprises the following steps:

[0177] Step 1: Arrange a 3×3 measurement grid in the target area, use the center point of each grid as a measurement point, and collect the response voltage component of each measurement point;

[0178] Step 2: Calculate the magnetic gradient tensor of each measurement point based on the response voltage component of each measurement point, calculate the target position vector of each measurement point based on the magnetic gradient tensor and the response voltage component, establish a global coordinate system in the measurement grid, generate the global target position of the measurement point based on the target position vectors of all measurement points, and average the global target positions of all measurement points to generate an average global target position;

[0179] Step 3: Construct a single dipole model at each measurement point, equating the underground target to a magnetic dipole with a dipole moment, and predict the response voltage component using the single dipole model;

[0180] Step 4: Use the average global target position as the initial value of the Gauss-Newton algorithm, construct a residual matrix that reflects the difference between the actual response voltage and the response voltage predicted by the single dipole model, take the negative derivative of the residual matrix to generate the Jacobian matrix, and iterate from the initial value to update the average global target position until the maximum number of iterations is reached. Output the last updated average global target position as the positioning result;

[0181] Step 5: Obtain and decompose the magnetic gradient tensor in the area where the positioning result is located, generate the target aspect ratio, and simultaneously measure the soil conductivity. Use the soil conductivity to correct the decay time coefficient of the transient electromagnetic response to generate the actual decay time coefficient, which reflects the conductivity of the target.

[0182] Step 6: Collect the aspect ratios and actual decay time coefficients of known targets appearing near the surface in historical data, perform Min-Max normalization on the aspect ratios and actual decay time coefficients, generate the feature vector of each known target, calculate the Euclidean distance between all feature vectors of known targets, take the 30% quantile of the Euclidean distance as the neighborhood radius, and set the minimum number of samples for clustering. Perform unsupervised cluster analysis on the feature vector of each known target based on the DBSCAN clustering algorithm. The feature vector of each known target in the cluster is equivalent to a sample point. In each cluster, the mean of the sample points is calculated as the cluster center. The Mahalanobis distance from each sample point in the cluster to the cluster center is generated, and a Mahalanobis distance threshold is set. When the Mahalanobis distance from a sample point to the cluster center exceeds the Mahalanobis distance threshold, the sample point is removed from the cluster. After each group of sample points is removed, the cluster center is recalculated, and the Mahalanobis distances of other sample points to the cluster center are calculated. Repeat the above steps of removing sample points until there are no sample points in the cluster whose Mahalanobis distance from the cluster center exceeds the Mahalanobis distance threshold.

[0183] Step 7: Build a deep learning network, take the known target feature vectors in different clusters as input, and the category of the known target as the label, train the target recognition model for each cluster, calculate the feature vector of the real-time located target according to the above steps, select the cluster center with the closest Euclidean distance to the real-time located target feature vector, assign the real-time located target feature vector to the cluster corresponding to the cluster center, and input the real-time located target feature vector into the target recognition model of the corresponding cluster to identify the category of the real-time located target.

[0184] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0185] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software depends on the specific application and design constraints of the technical solution.

[0186] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0187] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A transient electromagnetic near-surface target feature recognition system, characterized in that: Specifically include: A data acquisition module is used to arrange a 3×3 measurement grid in the target area, use the center point of each grid as a measurement point, and collect the response voltage component of each measurement point; an initial position measurement module, configured to calculate a magnetic gradient tensor of each measurement point based on a response voltage component of each measurement point, calculate a target position vector of each measurement point based on the magnetic gradient tensor and the response voltage component, establish a global coordinate system in the measurement grid, generate a global target position of the measurement point based on the target position vectors of all measurement points, and average the global target positions of all measurement points to generate an average global target position; A model building module is used to construct a single dipole model at each measurement point, equating the underground target to a magnetic dipole with a dipole moment, and predicting the response voltage component through the single dipole model; The initial position optimization and positioning module is used to use the average global target position as the initial value of the Gauss-Newton algorithm, construct a residual matrix reflecting the difference between the actual response voltage and the response voltage predicted by the single dipole model, calculate the negative derivative of the residual matrix to generate the Jacobian matrix, iterate from the initial value to update the average global target position until the maximum number of iterations is reached, and output the last updated average global target position as the positioning result; The target feature extraction module is used to obtain and decompose the magnetic gradient tensor in the area where the positioning result is located, generate the target aspect ratio, and simultaneously measure the soil conductivity. The decay time coefficient of the transient electromagnetic response is corrected by the soil conductivity to generate the actual decay time coefficient, which reflects the conductivity of the target. The target clustering module is used to collect the aspect ratios and actual decay time coefficients of known targets appearing near the surface in historical data, perform Min-Max normalization on the aspect ratios and actual decay time coefficients, generate the feature vector of each known target, calculate the Euclidean distance between all known target feature vectors, take the 30% quantile of the Euclidean distance as the neighborhood radius, and set the minimum number of samples for clustering. Unsupervised cluster analysis is performed on the feature vector of each known target based on the DBSCAN clustering algorithm. The feature vector of each known target in the cluster is equivalent to a sample point. The mean of the sample points in each cluster is calculated as the cluster center, and the Mahalanobis distance from each sample point in the cluster to the cluster center is generated. A Mahalanobis distance threshold is set. When the Mahalanobis distance from a sample point to the cluster center exceeds the Mahalanobis distance threshold, the sample point is removed from the cluster. After each group of sample points is removed, the cluster center is recalculated, and the Mahalanobis distances of other sample points to the cluster center are calculated. The above steps of removing sample points are repeated until there are no sample points in the cluster whose Mahalanobis distance from the cluster center exceeds the Mahalanobis distance threshold. The target recognition module is used to build a deep learning network. It takes the known target feature vectors in different clusters as input and the category of the known target as the label. It trains the target recognition model for each cluster, calculates the feature vector of the real-time located target according to the above steps, selects the cluster center with the closest Euclidean distance to the real-time located target feature vector, assigns the real-time located target feature vector to the cluster corresponding to the cluster center, and inputs the real-time located target feature vector into the target recognition model of the corresponding cluster to identify the category of the real-time located target.

2. A transient electromagnetic near-surface target feature recognition system according to claim 1, characterized in that: In the data acquisition module, a JPMD sensor including a single-layer transmitting coil and five three-axis receiving coils is used for data acquisition. The spacing of the measurement grid is 25 cm, and 15 response voltage components are collected at each measurement point.

3. The transient electromagnetic near-surface target feature recognition system according to claim 2, characterized in that: In the JPMD sensor, the center of the JPMD sensor is used as the origin of the coordinate system. There is a three-axis receiving coil at the origin of the coordinate system, and the other four three-axis receiving coils are distributed in the four positive directions of the coordinate system origin, namely, front, back, left, and right. The positive direction of the x-axis is the direction pointing to the right of the coordinate system origin, the positive direction of the y-axis is the direction pointing to the back of the coordinate system origin, and the positive direction of the z-axis is perpendicular to the x-axis and y-axis, passes through the origin of the coordinate system, and points upward. A measurement point coordinate system is established, and the three-axis receiving coil at the origin of the coordinate system is set as R3, the coils in the positive and negative directions of the x-axis are set as R2 and R4, and the coils in the positive and negative directions of the y-axis are set as R1 and R5, respectively.

4. The transient electromagnetic near-surface target feature recognition system according to claim 3, characterized in that: In the initial position measurement module, the principle for calculating the target position vector of each measurement point is: The magnetic gradient tensor at the measurement point is: Where G represents the magnetic gradient tensor of the i-th measurement point, i represents the index of the measurement point, and i∈[1,9], g xx,i Indicates the rate of change of the magnetic field in the x direction at the i-th measurement point, g xy,i Indicates the rate of change of the magnetic field in the y direction at the i-th measurement point in the x direction, g xz,i Indicates the rate of change of the magnetic field in the z direction in the x direction at the i-th measurement point, g yx,i Indicates the rate of change of the magnetic field in the x-direction to the y-direction at the i-th measurement point, g yy,i Indicates the rate of change of the magnetic field in the y direction at the i-th measurement point, g yz,i Indicates the rate of change of the magnetic field in the z direction in the y direction at the i-th measurement point, g zx,i Indicates the rate of change of the magnetic field in the x-direction in the z-direction at the i-th measurement point, g zy,i Indicates the rate of change of the magnetic field in the y direction in the z direction at the i-th measurement point, g zz,i represents the rate of change of the magnetic field in the z direction at the i-th measurement point; g zz,i =-(g xx,i +g yy,i ) Among them, V 4x,i represents the x-axis response voltage component of R4 at the i-th measurement point, V 2x,i represents the x-axis response voltage component of R2 at the i-th measurement point, d represents the distance from R1 to the origin of the coordinate system, V 4y,i represents the y-axis response voltage component of R4 at the i-th measurement point, V 2y,i Represents the y-axis response voltage component of R2 at the i-th measurement point, V 4z,i represents the z-axis response voltage component of R4 at the i-th measurement point, V 2z,i represents the z-axis response voltage component of R2 at the i-th measurement point, V 1y,i Represents the y-axis response voltage component of R1 at the i-th measurement point, V 5y,i Represents the y-axis response voltage component of R5 at the i-th measurement point, V 1z,i represents the z-axis response voltage component of R1 at the i-th measurement point, V 5z,i represents the z-axis response voltage component of R5 in the i-th measurement point; The formula for calculating the target position vector of each measurement point is: Among them, R i Represents the target position vector of the i-th measurement point, V i Represents the response voltage component of the i-th measurement point.

5. The transient electromagnetic near-surface target feature recognition system according to claim 4, characterized in that: The principle behind generating the average global target position is: The center point of the 3×3 measurement grid is used as the global origin. On the plane where the measurement grid is located, there is a measurement point in each of the eight positive directions of the global origin: up, down, left, right, upper left, lower left, upper right, and lower right. The global x-axis is defined as the positive direction of the global origin pointing due right, the global y-axis is defined as the positive direction of the global origin pointing due upward, and the global z-axis is defined as the positive direction passing through the global origin, perpendicular to the global x-axis and global y-axis, and pointing outside the grid plane. The global coordinate system of the target area is constructed. For each measurement point, there is a known global coordinate and target position vector, then the global target position of a single measurement point is: r i =S i +R i Among them, r i represents the global target position of the i-th measurement point, S i represents the global coordinates of the i-th measurement point; The average global target position is: Among them, r d represents the average global target position.

6. The transient electromagnetic near-surface target feature recognition system according to claim 1, characterized in that: The principle for updating the target position in the initial position optimization and positioning module is as follows: The residual matrix is: F(r d )=V obs -V pre Among them, F(r d ) represents the residual matrix, V obs Represents the response voltage at the measurement point, V pre represents the response voltage predicted by the single-dipole model; The Jacobian matrix is: Where J represents the Jacobian matrix; The iterative equation for the target position is: r d k+1 =r d k +Δr d k Δr d k =(J T J) -1 J T F(r d k ) Among them, r d k+1 represents the average global target position at the k+1th iteration, r d k represents the average global target position of the kth iteration, k represents the index of the number of iterations, and k∈[0,K], K represents the maximum number of iterations. When k=0, r d 0 =r d , r d represents the initial average global target position before iteration, Δr d k Represents the correction amount to the average global target position at the kth iteration.

7. The transient electromagnetic near-surface target feature recognition system according to claim 1, characterized in that: The principle for generating the target aspect ratio in the target feature extraction module is as follows: The magnetic gradient tensor in the area where the positioning result is located is: Among them, G0 represents the magnetic gradient tensor of the area where the positioning result is located, B x 、B y 、B z Represent the magnetic field components of the x, y, and z axes respectively; The formula for decomposing G0 to obtain three eigenvalues ​​is: Solve the equation |G0-λ i I|=0, we get three eigenvalues ​​λ1, λ2, and λ3, and |λ1|>|λ2|>|λ3|, where I represents the identity matrix; The aspect ratio is: Where L represents the target aspect ratio.

8. The transient electromagnetic near-surface target feature recognition system according to claim 1, characterized in that: The principle of judging the conductivity of the target based on the corrected decay time coefficient in the target feature extraction module is as follows: The formula for correcting the decay time coefficient is: k=k0(1+kσ s ) Among them, k represents the actual decay time coefficient, κ0 represents the decay time coefficient before correction, k represents the empirical coefficient, σ s Indicates soil electrical conductivity; κ is proportional to the target conductivity.

9. The transient electromagnetic near-surface target feature recognition system according to claim 1, characterized in that: In the target clustering module, the principle of unsupervised cluster analysis is as follows: The formula for calculating the Euclidean distance between all known target feature vectors is: Among them, v o 、v p denote the oth and pth known target feature vectors respectively, d(v o ,v p ) means v o 、v p The Euclidean distance between them, o, p∈[1,P] and o≠p, P represents the total number of known target feature vectors, L ′ o represents the normalized aspect ratio of the oth known target feature vector, L′ p represents the normalized aspect ratio of the pth known target feature vector, k ′ o represents the normalized actual decay time coefficient of the oth known target feature vector, κ′ p represents the normalized actual decay time coefficient of the p-th known target feature vector; The neighborhood radius of the cluster is taken as the 30% quantile of the Euclidean distance of all sample points, and the minimum number of sample points for the cluster is Indicates taking the square root of the total number of known target feature vectors and then rounding up; randomly select a sample point and count the other sample points within the neighborhood radius. If the number exceeds The selected sample point is regarded as the current cluster center, and all sample points are assigned to the cluster where the current cluster center with the closest Euclidean distance is located, which is considered to complete the unsupervised cluster analysis; The cluster center is calculated in each cluster according to the formula: Among them, c q represents the cluster center of the qth cluster, q represents the cluster index, P q represents the number of sample points in the qth cluster, j represents the index of the sample point in the cluster, and j∈[1,J q ], J q represents the number of sample points in the qth cluster, v j Represents the target feature vector corresponding to the j-th sample point; The formula for generating the Mahalanobis distance from each eigenvector in a cluster to the cluster center is: Among them, D j,q S represents the Mahalanobis distance from the jth sample point to the cluster center in the qth cluster. q represents the intra-class covariance matrix of the qth cluster; When D j,q >D max When , the sample point is removed from the cluster, D max Represents the Mahalanobis distance threshold.

10. A transient electromagnetic near-surface target feature recognition method, characterized by: The method is performed by the transient electromagnetic near-surface target feature recognition system according to any one of claims 1 to 9, and the specific steps include: Step 1: Arrange a 3×3 measurement grid in the target area, use the center point of each grid as a measurement point, and collect the response voltage component of each measurement point; Step 2: Calculate the magnetic gradient tensor of each measurement point based on the response voltage component of each measurement point, calculate the target position vector of each measurement point based on the magnetic gradient tensor and the response voltage component, establish a global coordinate system in the measurement grid, generate the global target position of the measurement point based on the target position vectors of all measurement points, and average the global target positions of all measurement points to generate an average global target position; Step 3: Construct a single dipole model at each measurement point, equating the underground target to a magnetic dipole with a dipole moment, and predict the response voltage component using the single dipole model; Step 4: Use the average global target position as the initial value of the Gauss-Newton algorithm, construct a residual matrix that reflects the difference between the actual response voltage and the response voltage predicted by the single dipole model, take the negative derivative of the residual matrix to generate the Jacobian matrix, and iterate from the initial value to update the average global target position until the maximum number of iterations is reached. Output the last updated average global target position as the positioning result; Step 5: Obtain and decompose the magnetic gradient tensor in the area where the positioning result is located, generate the target aspect ratio, and simultaneously measure the soil conductivity. Use the soil conductivity to correct the decay time coefficient of the transient electromagnetic response to generate the actual decay time coefficient, which reflects the conductivity of the target. Step 6: Collect the aspect ratios and actual decay time coefficients of known targets appearing near the surface in historical data, perform Min-Max normalization on the aspect ratios and actual decay time coefficients, generate the feature vector of each known target, calculate the Euclidean distance between all feature vectors of known targets, take the 30% quantile of the Euclidean distance as the neighborhood radius, and set the minimum number of samples for clustering. Perform unsupervised cluster analysis on the feature vector of each known target based on the DBSCAN clustering algorithm. The feature vector of each known target in the cluster is equivalent to a sample point. In each cluster, the mean of the sample points is calculated as the cluster center. The Mahalanobis distance from each sample point in the cluster to the cluster center is generated, and a Mahalanobis distance threshold is set. When the Mahalanobis distance from a sample point to the cluster center exceeds the Mahalanobis distance threshold, the sample point is removed from the cluster. After each group of sample points is removed, the cluster center is recalculated, and the Mahalanobis distances of other sample points to the cluster center are calculated. Repeat the above steps of removing sample points until there are no sample points in the cluster whose Mahalanobis distance from the cluster center exceeds the Mahalanobis distance threshold. Step 7: Build a deep learning network, take the known target feature vectors in different clusters as input, and the category of the known target as the label, train the target recognition model for each cluster, calculate the feature vector of the real-time located target according to the above steps, select the cluster center with the closest Euclidean distance to the real-time located target feature vector, assign the real-time located target feature vector to the cluster corresponding to the cluster center, and input the real-time located target feature vector into the target recognition model of the corresponding cluster to identify the category of the real-time located target.

Images