A geomagnetic reference map construction method based on a super-resolution network

Abstract

The application belongs to the technical field of geomagnetic navigation, and particularly relates to a geomagnetic reference map construction method based on a super-resolution network, which comprises the following steps: S1: using a geomagnetic collection device to collect geomagnetic data points in a rectangular region to form sparse collection data, and constructing a sparse matrix as a geomagnetic test set; S2: obtaining geomagnetic library data from a website of a geological exploration bureau to obtain a dense matrix, performing downsampling on the dense matrix to obtain a sparse matrix, and combining the two to construct a geomagnetic training set; S3: inputting the geomagnetic training set into a super-resolution network composed of a generator and a discriminator to perform training; and S4: inputting the geomagnetic test set into the generator to obtain an output result, and further constructing a high-precision geomagnetic reference map. The geomagnetic reference map construction system based on the super-resolution network can more flexibly adapt to data of different scales and complexities, improves the accuracy of constructing the geomagnetic reference map, has a smaller mean square error compared with a traditional Kriging interpolation method, and has higher precision.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of geomagnetic navigation, and particularly relates to a geomagnetic reference map construction method based on a super-resolution network. BACKGROUND

[0002] In the past few decades, positioning technology has made significant progress, and inertial navigation technology and satellite positioning technology are the main research directions. However, the gyroscope used in inertial navigation has the problem of cumulative drift error over time, and satellite signals are easily affected by external factors such as terrain and climate, making it difficult to further improve the navigation and positioning accuracy. Considering that the geomagnetic field is relatively stable and less affected by environmental factors, geomagnetic navigation technology has gradually become an effective auxiliary positioning method, and constructing a high-precision geomagnetic reference map is a key step to realize accurate and stable navigation.

[0003] In recent years, the construction of the geomagnetic reference map mainly adopts the interpolation method, and common methods include the inverse distance weighted interpolation method, the radial basis function method and the Kriging interpolation method based on the particle swarm algorithm. Although these algorithms have improved the root mean square error evaluation index in the process of constructing the geomagnetic reference map, it is still difficult to accurately construct the detailed information between the magnetic measurement points, and therefore the application provides a geomagnetic reference map construction method based on a super-resolution network to solve the above problems. SUMMARY

[0004] (I) Technical problems solved

[0005] In view of the deficiencies of the prior art, the application provides a geomagnetic reference map construction method based on a super-resolution network, which solves the problems proposed in the background.

[0006] (II) Technical solutions

[0007] In order to achieve the above purpose, the application specifically adopts the following technical solutions:

[0008] A geomagnetic reference map construction method based on a super-resolution network comprises the following steps:

[0009] S1: using a geomagnetic collection device to collect geomagnetic data points in a rectangular region to form sparse collection data, and constructing a sparse matrix as a geomagnetic test set;

[0010] S2: obtaining geomagnetic database data from a geological exploration bureau website, extracting geomagnetic data points in a plurality of rectangular regions with the same size to form dense matrices, and performing bicubic interpolation down-sampling on the matrices to obtain sparse matrices. The dense matrices and the sparse matrices are combined to construct a geomagnetic training set;

[0011] S3: inputting the sparse matrix I LRThe input is fed into a generator consisting of a feature extraction module, a feature enhancement module, and an upsampling module to obtain a dense matrix I. SR Then I SR With the dense matrix I in the original geomagnetic training set HR The input is fed into a discriminator composed of convolutional layers, deep convolutional modules, and activation functions for judgment. If the dense matrix generated by the generator is I... HR If the weights are correct, the weights are updated and training continues; otherwise, training ends. During this period, the generator and discriminator are trained alternately to minimize the loss function, causing the generator to generate a dense matrix.

[0012] S4: Input the sparse matrix composed of geomagnetic data points in the geomagnetic test set into the trained generator to obtain a dense matrix, and prepare a high-precision geomagnetic reference map corresponding to the collection area.

[0013] Furthermore, the geomagnetic data in the geomagnetic test set in S1 comes from a geomagnetic data acquisition device, using a Wirt WT10F gaussmeter with a sensitivity of 0.0001 Gs, suitable for measuring the magnetic field strength of the Earth's environment. A specific open rectangular area is selected for magnetic field strength measurement; outliers are removed during the measurement process, and measurements are repeated. The average value of three repeated measurements is ultimately selected as the sample point. The collected sample points include the three components B of the magnetic field. x B y B z The expression for the total magnetic field strength is: Geomagnetic data points of size 64×64 were extracted to form a sparse matrix as a geomagnetic test set.

[0014] Furthermore, the geomagnetic data in the geomagnetic training set in S2 comes from the website of the Geological Exploration Bureau. Several rectangular magnetic field value points of size 128×128 are extracted to form a matrix. The matrix is ​​downsampled using bicubic interpolation to obtain a matrix of size 64×64. The two are combined to construct the geomagnetic training set.

[0015] Furthermore, the bicubic interpolation formula is: In the formula, I(x+i,y+j) represents the magnetic field value matrix of the original data at different latitudes and longitudes; ω i,j The weights are calculated based on the bicubic interpolation kernel function; P(x,y) is the magnetic field value matrix at different latitudes and longitudes after downsampling.

[0016] Furthermore, the specific content of inputting the geomagnetic training set into the generator and discriminator for training in S3 is as follows: inputting the sparse matrix I in the geomagnetic training set... LR The input is fed into the Conv layer of the feature extraction module in the generator G for feature extraction, resulting in the output feature matrix F1, expressed as: F1 = Conv(ILR ,ω1)+b1, where ω1 is the convolution kernel weight and b1 is the bias term. The convolution kernel size is set to 3×3 and the stride is 1. The PReLU activation function in the feature extraction module is used to perform non-linear mapping on the output F1 of the convolution layer to obtain the activated feature matrix F2, which is expressed as: F2=PReLU(F1).

[0017] The obtained feature matrix F2 is fed into a feature enhancement module consisting of two Conv layers, two PReLU layers, one BN layer, and one Sum layer. This process is repeated four times to efficiently learn complex features and stabilize the learning process, outputting a feature matrix F3. The convolution kernel size is set to 3×3, and the stride is 1. A Conv layer, PReLU layer, and Sum layer are added between the feature enhancement module and the upsampling module to process feature matrix F3, resulting in feature matrix F6. The expression is: F6 = F5 + I LR F4 = Conv(F3,ω1) + b1, F5 = PReLU(F4), the convolution kernel size is set to 1×1 to preserve the input feature information, increase the non-linear expression of the feature matrix, and accelerate the learning of the network through residual connections to ensure that information is not lost during the upsampling process;

[0018] F6 is input into an upsampling module consisting of a Conv layer, a PixelShuffer layer, and a PReLU layer, and this process is repeated. Finally, the output result I is obtained by passing the input through a 9×9 convolution kernel. SR ;

[0019] The output result I in the generator SR The matrix I composed of the original dense geomagnetic data points HR The inputs are fed into the Conv layer of the discriminator D for convolution processing. The convolution kernel is 1×1. Then, the resulting feature matrix is ​​activated by the LeakyReLU function to output a non-linear feature matrix.

[0020] The obtained nonlinear feature matrix is ​​input into a deep convolutional module consisting of Conv layer, BN layer and LeakyReLU layer for processing, and the number of times is set to 7. The process is repeated, and the convolutional kernel is 3×3.

[0021] Construct a dense fully connected layer, pass it through a LeakyReLU activation function, then pass it through another dense fully connected layer, and apply the result to a sigmoid function for non-linear activation; the output is 0 or 1, used to determine if the dense matrix generated by the generator is I. HR Or I SR ;

[0022] The loss function expression for the generator is: Where L G D represents the adversarial loss function, D(·) represents the discriminant mapping function, G(·) represents the generative mapping function, and E represents the expectation. Representing I LR It follows a P-distribution.

[0023] The discriminator's loss function is an adversarial loss function, expressed as:

[0024]

[0025] Where L D D represents the adversarial loss function, D(·) represents the discriminant mapping function, G(·) represents the generative mapping function, and E represents the expectation. I represents the distribution of the training data. HR ~P train (I HR ) represents I HR Follows a P-distribution. This represents the data distribution after passing through the generator. Representing I LR It follows a P-distribution.

[0026] Furthermore, the parameter-corrected linear unit of the PReLU activation function is an improved version of ReLU, and the mathematical expression of the PReLU activation function is:

[0027]

[0028] Among them, a i It is a constant, x i It is the input, y i This is the output.

[0029] Furthermore, the LeakyReLU activation function is an improved version of ReLU, introducing a negative non-zero gradient, and its mathematical expression is:

[0030]

[0031] Where, a i It is a fixed parameter in the interval (1, +∞), x i It is the input, y i This is the output result.

[0032] Furthermore, the mathematical expression for the Sigmoid function is:

[0033]

[0034] Where x represents the input and y represents the output.

[0035] (III) Beneficial Effects

[0036] Compared with existing technologies, this invention provides a method for constructing geomagnetic reference maps based on super-resolution networks, which has the following beneficial effects:

[0037] Traditional Kriging interpolation methods suffer from parameter selection issues when processing data from regions of varying sizes. The geomagnetic benchmark map construction system based on super-resolution networks can more flexibly adapt to data of different scales and complexities, improving the accuracy of geomagnetic benchmark map construction. Furthermore, super-resolution networks can better capture detailed information between magnetic measurement points, including local features and subtle changes, thus achieving more accurate geomagnetic benchmark map construction with a smaller mean square error compared to traditional Kriging interpolation. Attached Figure Description

[0038] Figure 1 This is a flowchart of the method steps of the present invention;

[0039] Figure 2 This is a block diagram of the geomagnetic data acquisition system of the present invention;

[0040] Figure 3 This is a schematic diagram of the generator structure of the present invention;

[0041] Figure 4 This is a schematic diagram of the discriminator structure of the present invention;

[0042] Figure 5 This is a schematic diagram of the feature extraction module structure in the generator of the present invention;

[0043] Figure 6 This is a schematic diagram of the feature enhancement module structure in the generator of the present invention;

[0044] Figure 7 This is a schematic diagram of the upsampling module structure in the generator of this invention;

[0045] Figure 8 This is a schematic diagram of the depthwise convolution module structure in the discriminator of the present invention;

[0046] Figure 9 This is a schematic diagram comparing the method proposed in this invention with the Kriging interpolation method. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0048] Example

[0049] like Figures 1-9 As shown, a geomagnetic reference map construction system based on super-resolution networks includes the following steps:

[0050] S1: Use geomagnetic acquisition equipment to collect geomagnetic data points in a rectangular area, form sparse acquisition data, and construct a sparse matrix as a geomagnetic test set.

[0051] The geomagnetic data used in the geomagnetic test set comes from sources such as Figure 2 The geomagnetic data acquisition device shown uses a Wheatell WT10F gaussmeter with a sensitivity of 0.0001 Gs (approximately 10 nT), which is suitable for measuring the magnetic field strength of the Earth's environment. Figure 2 The magnetic sensor and conditioning circuit shown are used for the detection and conditioning of geomagnetic component signals; the A / D conversion module is responsible for real-time acquisition of analog signals and digital conversion; the memory module is used for real-time signal storage; the FPGA controls the working mode of the acquisition, storage, and communication modules, coordinating the operation of each module; the communication interface module realizes data feedback and transmission with the host computer; the power supply module converts the input power into the voltage required by each module and provides power support. A rectangular open area was selected for magnetic field strength measurement. Outliers were eliminated during the measurement process, and measurements were repeated. The average value of three repeated measurements was finally selected as the sample point. The collected sample points include the three components of the magnetic field, B. x B y B z The expression for the total magnetic field strength is: Furthermore, geomagnetic data points of size 64×64 are extracted to form a sparse matrix as a geomagnetic test set.

[0052] S2: Obtain geomagnetic database data from the geological exploration bureau's website, extract geomagnetic data points from several rectangular regions of the same size to form a dense matrix, and perform bicubic interpolation downsampling on these matrices to obtain a sparse matrix. Combine the above dense and sparse matrices to construct a geomagnetic training set.

[0053] The geomagnetic data used in the geomagnetic training set of this invention comes from a geological exploration bureau. A matrix is ​​formed by extracting magnetic field value points from several rectangular regions of size 128×128. This matrix is ​​downsampled using bicubic interpolation to obtain a 64×64 matrix, and the two are combined to construct the training set. The bicubic interpolation formula is as follows: In the formula, I(x+i,y+j) represents the magnetic field value matrix of the original data at different latitudes and longitudes; ω i,j The weights are calculated based on the bicubic interpolation kernel function; P(x,y) is the magnetic field value matrix at different latitudes and longitudes after downsampling.

[0054] S3: The sparse matrix I in the geomagnetic training setLR The input is fed into a generator consisting of a feature extraction module, a feature enhancement module, and an upsampling module to obtain a dense matrix I. SR Then I SR With the dense matrix I in the original geomagnetic training set HR The input is fed into a discriminator composed of convolutional layers, deep convolutional modules, and activation functions for judgment. If the dense matrix generated by the generator is I... HR If the weights are correct, the weights are updated and training continues; otherwise, training ends. During this period, the generator and discriminator are trained alternately to minimize the loss function, causing the generator to generate a dense matrix.

[0055] The sparse matrix I in the geomagnetic training set LR Input to such Figure 3 The Conv layer in the feature extraction module of the generator G shown in the figure extracts features to obtain the output feature matrix F1, which is expressed as: F1 = Conv(I LR ,ω1)+b1, where ω1 is the kernel weight and b1 is the bias term. The kernel size is set to 3×3, and the stride is 1. The output F1 of the convolutional layer is applied as follows: Figure 5 The PReLU activation function in the feature extraction module performs a non-linear mapping to obtain the activated feature matrix F2, expressed as: F2 = PReLU(F1). The parameter-corrected linear unit of the PReLU activation function is an improved version of ReLU, and the mathematical expression of the PReLU activation function is:

[0056]

[0057] Among them, a i It is a very small constant, x i It is the input, y i This is the output.

[0058] The obtained feature matrix F2 is fed into a system consisting of two Conv layers, two PReLU layers, one BN layer, and one sum layer, as shown in the figure. Figure 6 The feature enhancement module shown is used, and the number of iterations is set to 4. This process is repeated to efficiently learn complex features and stabilize the learning process. The output is the feature matrix F3, where the convolution kernel size is set to 3×3 and the stride is 1.

[0059] A Conv layer, a PReLU layer, and a Sum layer are added between the feature enhancement module and the upsampling module to process the feature matrix F3, resulting in the feature matrix F6, expressed as: F6 = F5 + I LRF4 = Conv(F3,ω1) + b1, F5 = PReLU(F4), where the convolution kernel size is set to 1×1 to preserve the input feature information, increase the non-linear expression of the feature matrix, and accelerate the learning of the network through residual connections to ensure that information is not lost during the upsampling process.

[0060] Input F6 into a system consisting of a Conv layer, a PixelShuffer layer, and a PReLU layer, as shown below. Figure 7 The upsampling module shown is repeated, and finally the output result I is obtained through a 9×9 convolution kernel. SR .

[0061] The output result I in the generator SR The matrix I composed of dense geomagnetic data points in the original geomagnetic training set HR Enter them together as follows Figure 4 The discriminator D shown in the diagram performs convolutional processing on the Conv layer with a 1×1 kernel. The resulting feature matrix is ​​then passed through the LeakyReLU activation function, outputting a non-linear feature matrix. The LeakyReLU activation function is an improved version of ReLU, introducing negative non-zero gradients. Its mathematical expression is:

[0062]

[0063] Where, a i It is a fixed parameter in the interval (1, +∞), x i It is the input, y i This is the output result.

[0064] The obtained nonlinear characteristic matrix is ​​input into a system consisting of Conv layers, BN layers, and LeakyReLU layers, as shown in the figure. Figure 8 The process is performed in the depthwise convolution module shown, and this process is repeated 7 times. The convolution kernels are all 3×3.

[0065] The output from the previous step is fed into a dense fully connected layer and a LeakyReLU layer, followed by another dense fully connected layer. This introduces non-linearity and complexity, helping the discriminator learn more complex and abstract features, thus improving its ability to distinguish between real and generated data. The result is then non-linearly activated using the sigmoid function; the output is either 0 or 1, used to determine whether the input matrix comes from the generator network or the geomagnetic training set. The mathematical expression for the sigmoid function is:

[0066]

[0067] Where x represents the input and y represents the output.

[0068] The loss function expression for the generator is: Where L G D represents the adversarial loss function, D(·) represents the discriminant mapping function, G(·) represents the generative mapping function, and E represents the expectation. Representing I LR It follows a P-distribution.

[0069] The discriminator's loss function is an adversarial loss function, expressed as:

[0070]

[0071] Where L D D represents the adversarial loss function, D(·) represents the discriminant mapping function, G(·) represents the generative mapping function, and E represents the expectation. I represents the distribution of the training data. HR ~P train (I HR ) represents I HR Follows a P-distribution. This represents the data distribution after passing through the generator. Representing I LR It follows a P-distribution.

[0072] The discriminator undergoes 1000 iterations, and the generator also undergoes 1000 iterations. The generator and discriminator are trained alternately to minimize the loss function, so that the generator generates a dense matrix.

[0073] S4: Input the sparse matrix composed of geomagnetic data points in the geomagnetic test set into the trained generator to obtain a dense matrix, and prepare a high-precision geomagnetic reference map corresponding to the collection area.

[0074] The mean square error results of comparing the method in this invention with the Kriging interpolation method are shown in the table below, and the comparison graph is shown in the figure below. Figure 9 As shown.

[0075]

[0076] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A geomagnetic reference map construction method based on a super-resolution network, characterized by comprising the following steps: It comprises the following steps: S1: collecting geomagnetic data points in a rectangular area using a geomagnetic collection device to form sparse collection data and construct a sparse matrix as a geomagnetic test set; S2: obtaining geomagnetic library data from a geological exploration bureau website, extracting geomagnetic data points in several rectangular areas of the same size to form dense matrices, and performing bicubic interpolation downsampling on the matrices to obtain sparse matrices; and combining the dense matrices and the sparse matrices to construct a geomagnetic training set; S3: sparse matrix I in the geomagnetic training set LR Input into the generator composed of feature extraction module, feature enhancement module and up-sampling module structure, get dense matrix I SR , I SR And the dense matrix I HR In the original geomagnetic training set, input into the discriminator composed of convolution layer, deep convolution module and activation function structure for judgment, if the dense matrix generated by the generator is I HR , then update the weight and continue training; otherwise, the training is ended; during this period, the generator and the discriminator are alternately trained to minimize the loss function, so that the generator generates dense matrix; S4: inputting the sparse matrix composed of geomagnetic data points in the geomagnetic test set into the trained generator to obtain a dense matrix, and preparing a high-precision geomagnetic reference map corresponding to the collection area.

2. The geomagnetic reference map construction method based on a super-resolution network according to claim 1, characterized in that: The geomagnetic data in the geomagnetic test set in S1 is sourced from a geomagnetic data collection device, which adopts a WT10F gauss meter with a sensitivity of 0.0001Gs and is suitable for measuring the magnetic field strength of the terrestrial environment; Select a certain open rectangular area to measure the magnetic field strength, eliminate the abnormal value in the measurement process and re-measure, finally select the average value of three repeated measurements as the sample point; The collected sample points include three components B x 、 B y 、 B z , the total magnetic field intensity expression is: And extract the size of 64*64 geomagnetic data points to form a sparse matrix as the geomagnetic test set. 3.The geomagnetic reference map construction method based on super-resolution network according to claim 1, characterized in that: The geomagnetic data in the geomagnetic training set in S2 is sourced from a geological exploration bureau website, and a plurality of rectangular area magnetic field value points with a size of 128x128 are extracted to form a matrix; bicubic interpolation is used to downsample the matrix to obtain a matrix with a size of 64x64, and the two matrices are combined to construct a geomagnetic training set.

4. The geomagnetic reference map construction method based on super-resolution network according to claim 3, characterized in that: The dual cubic interpolation formula is: In the formula, I(x+i, y+j) represents a magnetic field value matrix of the original data at different longitudes and latitudes; ω i,j is a weight calculated according to a dual cubic interpolation kernel function; and P(x, y) is a magnetic field value matrix at different longitudes and latitudes after down-sampling.

5. The geomagnetic reference map construction method based on super-resolution network according to claim 1, characterized in that: The specific content of inputting the geomagnetic training set in S3 into the generator and the discriminator for training is that the sparse matrix I in the geomagnetic training set is inputted into the generator G and the discriminator D for training. LR The Conv layer in the feature extraction module inputted into the generator G is used for feature extraction to obtain an output feature matrix F1, and the expression is F1=Conv(I LR ,ω1)+b1, wherein ω1 is a convolution kernel weight, b1 is a bias term, the convolution kernel size is set to 3×3, and the step is 1; the PReLU activation function in the feature extraction module is used for nonlinear mapping on the convolution layer output F1 to obtain an activated feature matrix F2, and the expression is F2=PReLU(F1). The obtained feature matrix F2 is input into a feature enhancement module composed of two Conv layers, two PReLU layers, one BN layer and one sum layer, and the number of times is set to 4, and the process is repeated to efficiently learn complex features and stabilize the learning process, and the output feature matrix F3 is obtained, wherein the convolution kernel size is set to 3x3 and the step is set to 1; A Conv layer, a PReLU layer and a Sum layer are added between the feature enhancement module and the up-sampling module to process the feature matrix F3 to obtain a feature matrix F6; an expression is F6=F5+I LR wherein F4=Conv(F3, ω1)+b1, F5=PReLU(F4), a convolution kernel size is set as 1*1 to maintain input feature information, increase nonlinear expression of the feature matrix, and accelerate learning of the network through residual connection, and ensure that information is not lost in the up-sampling process. F6 is input to an up-sampling module composed of a Conv layer, a PixelShuffer layer, and a PReLU layer, and this process is repeated, and finally an output result I is obtained through a 9x9 convolution kernel SR ; The output result I in the generator SR and the original dense magnetic data point matrix I HR are input into the Conv layer in the discriminator D for convolution processing, the convolution kernel is 1*1, and then the obtained feature matrix is subjected to a LeakyReLU activation function, and a nonlinear feature matrix is output. The obtained nonlinear feature matrix is input into a deep convolution module composed of a Conv layer, a BN layer and a LeakyReLU layer for processing, and the number of times is set to 7, and the process is repeated, wherein the convolution kernel is 3x3; A denselayer fully connected layer is constructed, which passes through a LeakyReLU activation function and then passes through a denselayer fully connected layer, and the obtained result is nonlinearly activated by a sigmoid function; Output result 0 or 1, used to determine whether the dense matrix generated by the generator is I HR or I SR ; The loss function expression of the generator is: Wherein L G represents the adversarial loss function, D(·) represents the discriminant mapping function, G(·) represents the generation mapping function, E represents the expectation, represents I LR obeys P distribution; The loss function of the discriminator is an adversarial loss function, and the expression is: where L D represents an adversarial loss function, D(·) represents a discriminant mapping function, G(·) represents a generator mapping function, E represents an expectation, represents a distribution of training data, I HR ~ P train (I HR ) represents I HR obeys a P distribution, represents a distribution of data after passing through the generator, represents I LR obeys a P distribution.

6. The geomagnetic reference map construction method based on super-resolution network according to claim 5, characterized in that: The PReLU activation function is an improved version of the ReLU parameter correction unit, and the mathematical expression of the PReLU activation function is: where a i is a constant, x i is the input, and y i is the output.

7. The geomagnetic reference map construction method based on super-resolution network according to claim 5, characterized in that: The LeakyReLU activation function is an improved version of the Relu, which introduces a negative non-zero gradient, and its mathematical expression is: where a i is a fixed parameter in the interval (1, +∞), x i is the input, and y i is the output result.

8. The geomagnetic reference map construction method based on super-resolution network according to claim 5, characterized in that: The mathematical expression of the Sigmoid function is: Where x represents the input and y represents the output.

Images