A neural network-based aeromagnetic compensation method, device, and storage medium

Abstract

This invention provides a neural network-based aeromagnetic compensation method, device, and storage medium, relating to the field of airborne magnetic detection. The method includes: calibration flight; acquiring calibration flight data; selecting time points; processing the calibration flight data to construct an input group; constructing and dividing the dataset based on the input group; establishing an aeromagnetic compensation architecture based on a neural network, specifically including a geomagnetic network, an interference network, and a diurnal variation network; training the aeromagnetic compensation architecture using the dataset; retaining the interference network as the aeromagnetic compensation model after training; and normalizing the flight data to be processed before inputting it into the aeromagnetic compensation model to achieve aeromagnetic compensation. The device and storage medium are used to implement the method. The beneficial effects of this invention are: improved compensation effect in verification flight experiments; excellent aeromagnetic interference compensation effect even in complex magnetic measurement environments; and improved quality of airborne geophysical data.

Description

Technical Field

[0001] This invention relates to the field of airborne magnetic detection, and in particular to an airborne magnetic compensation method, device and storage medium based on neural networks. Background Technology

[0002] The spatial distribution of the Earth's magnetic field provides crucial information for solid and oil / gas resource exploration, unexploded ordnance detection, anti-submarine warfare, and underground ore body exploration. With rapid economic development and the continuous depletion of mineral resources, the nation's demand for underground resources is constantly increasing. As easily exploitable near-surface mineral resources are nearing depletion, exploration efforts are gradually shifting towards deep mineral deposits and mining in remote areas, placing high demands on existing exploration technologies. Airborne magnetic surveying, as an important method in geophysical exploration, is widely used in Earth science research, resource exploration, and military reconnaissance. During airborne magnetic surveys, the flight platform carrying the magnetic instruments cannot be completely demagnetized, resulting in magnetic interference that is mixed into the magnetic survey data during movement. This can cause the magnetic survey data to fail to accurately reflect the magnetic characteristics of the measured area, thus requiring compensation.

[0003] There are two types of aeromagnetic compensation methods: physical model-based methods achieve compensation by physically modeling the interfering magnetic field (such as the TL model) and solving the magnetic interference coefficient in the model. Essentially, it simplifies the nonlinear problem into a linear problem, which is simple to calculate, but has many problems such as multicollinearity. Non-physical model-based methods first establish a neural network magnetic compensation model, then obtain the interfering magnetic field network through iterative training, and finally obtain the output interfering magnetic field by measuring the corresponding input data during flight, thereby completing the compensation.

[0004] Current research on neural network magnetic compensation methods focuses on improving compensation accuracy and generalization performance by addressing both model structure and training process. Methods include introducing random noise for model regularization, using recurrent neural networks to enhance long-term generalization performance, and employing genetic algorithms to optimize initialization parameters. However, research on compensation in complex and challenging magnetic measurement scenarios has received relatively little attention. In such environments, existing neural network methods exhibit poor compensation performance and low computational efficiency. Summary of the Invention

[0005] The purpose of this invention is to address the problem of poor aeromagnetic interference compensation performance in complex magnetic measurement environments using existing technologies. This invention provides an aeromagnetic compensation method, device, and storage medium based on neural networks. The aeromagnetic compensation method based on neural networks mainly includes the following steps:

[0006] S1. Calibration Flight: Obtain calibration flight data.

[0007] S2. Select time points, process and calibrate flight data to construct input groups, and construct and divide datasets based on input groups.

[0008] The input group includes the total magnetic field value and the normalized results of time, direction cosine, and spatial variables obtained from the processed calibration flight data;

[0009] S3. Establish an aeromagnetic compensation architecture based on neural networks, specifically including geomagnetic networks, interference networks, and diurnal variation networks;

[0010] S4. Use the dataset to train the aeromagnetic compensation architecture, and retain the interference network as the aeromagnetic compensation model after training.

[0011] S5. Normalize the flight data to be processed and input it into the aeromagnetic compensation model to achieve aeromagnetic compensation.

[0012] Furthermore, calibration flight is the flight of an aircraft at a preset flight altitude and the range of flight altitude changes, during which small-amplitude maneuvers such as pitch, roll, and yaw are repeatedly performed.

[0013] The calibration flight data includes the aircraft's attitude, position, and time information, as well as measurements from the three-axis fluxgate magnetometer and the optical pump magnetometer.

[0014] Furthermore, step S2 specifically includes:

[0015] S21. Select time points during the calibration flight process according to preset time intervals;

[0016] S22. Process the calibration flight data corresponding to the time point, and construct the original variable set corresponding to the time point, specifically as follows:

[0017] S221. Combine the selected time point with the square of the calculated time point as the time variable;

[0018] S222. Combine the aircraft attitude information corresponding to the selected time point with the measurement value of the three-axis fluxgate magnetometer to calculate the direction cosine, and calculate and obtain the quadratic term and the differential term of the direction cosine, and combine them as the direction cosine variable.

[0019] S223. Select the longitude, latitude, and altitude from the aircraft's location information corresponding to the selected time point, and combine the calculated squared value of longitude, squared value of latitude, and the product of longitude and latitude as the location variable;

[0020] S224. Combine the time variable, direction cosine variable, and position variable at the selected time point as a set of original variables;

[0021] S23. Obtain the optical pump magnetometer measurement values ​​at time points as the total field magnetic force value;

[0022] S24. Normalize all data in the original variable groups according to the variable type, and then combine the total magnetic field value corresponding to the time point as the training group. Construct a dataset with the training group and divide it into training set and validation set.

[0023] Furthermore, the geomagnetic network is a one-dimensional convolutional neural network, consisting of convolutional layers, activation layers, pooling layers, flattening layers, and fully connected layers. The input is the normalized position variable, and the output is the fitted geomagnetic field value.

[0024] Furthermore, the interference network consists of a one-dimensional convolutional neural network and an ECA module. The input is a normalized direction cosine variable, and the output is the fitted interference magnetic field value.

[0025] Furthermore, the diurnal variation network consists of four fully connected layers. The first three fully connected layers have 20 neurons and use the tanh activation function, while the last fully connected layer has only one neuron and uses a linear activation function. The input is a normalized time variable, and the output is the fitted geomagnetic diurnal variation value.

[0026] Furthermore, the training process for the aeromagnetic compensation architecture is as follows:

[0027] Furthermore, step S5 specifically includes:

[0028] The flight data obtained during the test flight is normalized and then input into the flight compensation network to obtain the predicted magnetic interference value. The predicted magnetic interference value is then removed from the total field magnetic force value measured by the optical pump magnetometer to complete the aeromagnetic compensation operation.

[0029] A storage medium storing instructions and data for implementing a neural network-based aeromagnetic compensation method.

[0030] A computer device includes: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement a neural network-based aeromagnetic compensation method.

[0031] The beneficial effects of the technical solution provided by this invention are as follows: By refining the aeromagnetic compensation architecture based on neural networks and using a 1D CNN-Attention model in the interference network, this invention improves the generalization ability of the convolutional network by adding an ECA attention mechanism between convolutional layers, thereby improving the compensation effect of the verification flight experiment. It can also have a good aeromagnetic interference compensation effect in complex magnetic measurement environments, thus improving the quality of airborne geophysical data. Attached Figure Description

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0033] Figure 1 This is a flowchart of an aeromagnetic compensation method based on a neural network in an embodiment of the present invention;

[0034] Figure 2 This is a schematic diagram of the aeromagnetic compensation architecture in an embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram of the ECA module structure in an embodiment of the present invention;

[0036] Figure 4 This is a schematic diagram of global pooling in an embodiment of the present invention;

[0037] Figure 5 This is a schematic diagram of the hardware device working in an embodiment of the present invention. Detailed Implementation

[0038] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0039] Embodiments of the present invention provide a neural network-based aeromagnetic compensation method, device, and storage medium.

[0040] Please refer to Figure 1 , Figure 1 This is a flowchart of an aeromagnetic compensation method based on a neural network according to an embodiment of the present invention, which specifically includes the following steps:

[0041] The first step is calibration flight, which involves acquiring calibration flight data.

[0042] Calibration flight is the flight of an aircraft at a preset flight altitude and the range of altitude changes, during which small-amplitude maneuvers such as pitch, roll, and yaw are repeatedly performed.

[0043] The calibration flight data includes the aircraft's attitude, position, and time information, as well as measurements from the three-axis fluxgate magnetometer and the optical pump magnetometer.

[0044] The second step involves selecting time points, processing the calibration flight data, and constructing input groups. Data sets are then built and divided into units based on these input groups. Each input group contains the total magnetic field value and the normalized results of the time, direction cosine, and spatial variables obtained from processing the calibration flight data. Specifically:

[0045] Step 1: Select time points during the calibration flight process according to preset time intervals.

[0046] Step two involves processing the calibration flight data corresponding to the given time points and constructing the original variable set for each time point. Specifically:

[0047] (1) Combine the selected time point with the square of the calculated time point as the time variable;

[0048] (2) Calculate the direction cosine by combining the aircraft attitude information corresponding to the selected time point with the measurement value of the three-axis fluxgate magnetometer, and calculate and obtain the quadratic term and the differential term of the direction cosine, and combine them as the direction cosine variable.

[0049] (3) Select the longitude, latitude and altitude from the aircraft's location information corresponding to the selected time point, and combine the calculated square value of longitude, square value of latitude and the product of longitude and latitude as the location variable;

[0050] (4) Combine the time variable, direction cosine variable and position variable of the selected time point as a set of original variables.

[0051] Step 3: Obtain the optical pump magnetometer measurement value at the time point as the total field magnetic force value.

[0052] Step four: Normalize all data in the original variable groups according to the variable type, and then combine the total magnetic field value corresponding to the time point as the training group. Construct a dataset with the training group and divide it into training set and validation set.

[0053] The third step involves establishing an aeromagnetic compensation architecture based on neural networks, specifically including geomagnetic networks, interference networks, and diurnal variation networks, such as... Figure 2 As shown.

[0054] These three network designs are based on measurement characteristics. In airborne magnetic surveys, the data collected by the total field magnetometer includes various magnetic sources:

[0055]

[0056] In the formula, H Total-Mag These are measurements taken by the optical pump magnetometer. H G This is the value of the geomagnetic field. H Sun This is the diurnal variation of the geomagnetic field. H I This is the value of the interfering magnetic field.

[0057] The geomagnetic network is a one-dimensional convolutional neural network, consisting of convolutional layers, activation layers, pooling layers, flattening layers, and fully connected layers. The input is the normalized location variable, and the output is the fitted geomagnetic field value.

[0058] The formula for forward propagation in a one-dimensional convolutional layer is as follows:

[0059]

[0060] in, This is the input to the j-th neuron in the l-th layer. The bias of the j-th neuron in the l-th layer is given by [insert bias here]. This is the output of the i-th neuron in the (l-1)-th layer. It refers to the nuclei of the i-th neuron in layer (l-1) to the j-th neuron in layer l. This is used to perform one-dimensional convolution. In the geomagnetic network, the kernel size is set to 3, and the convolution layer mode is zero-padding. Convolution operations only begin when the center of the filter coincides with the corner of the input tensor. The input array... With output array They have the same dimensions. Activation function in the activation layer The intermediate output obtained for:

[0061]

[0062] In one-dimensional convolutional neural networks, this paper uses ReLU as the activation function. The formula for the ReLU activation function is as follows:

[0063]

[0064] Therefore, when input In the negative half-region, the intermediate output of the ReLU activation function The value is zero when the input is zero. In the positive half-zone, the intermediate output With input The values ​​are the same. The derivative of the ReLU function in the negative half-region is 0, at which point the output of the activation function is 0, and the neuron is not activated. The derivative of the ReLU function in the positive half-region is 1. Therefore, the ReLU function inhibits the activation of some neurons, making the neurons sparsely activated. In neural network compensation, the ReLU activation function is simple to calculate, but it can achieve excellent compensation results with the shortest training time. Using the tanh activation function will increase the training time and reduce the compensation effect. The compensation effect of using the Leaky ReLU activation function is close to that of the ReLU activation function, but it will significantly increase the time cost of the neural network compensation method.

[0065] The interference network consists of a one-dimensional convolutional neural network and an ECA module. The input is a normalized direction cosine variable, and the output is the fitted interference magnetic field value.

[0066] Since the interference network has 18 input terms and the causal relationship between input and output is relatively complex under complex magnetic measurement conditions, using a deeper convolutional network in the interference network can achieve better nonlinear fitting ability by superimposing nonlinear functions, thereby better learning the nonlinear characteristics of the interference network.

[0067] An ECA module is used between convolutional layers to calculate the attention weights corresponding to the interference magnetic field characteristics of each channel in the convolutional layer. The structure of the ECA module is as follows: Figure 3 As shown, a 3*1 convolution and ReLU function are used to extract the feature information of magnetic interference contained in the input. An ECA module is used between the convolutional layers to calculate the attention weights corresponding to the interference magnetic field features of each channel in the convolutional layer. This makes the interference magnetic field features of different channels have different effects on the output of the interference network.

[0068] The ECA module receives the interfering magnetic field features input from the convolutional layer and processes them through global average pooling, as follows: Figure 4 As shown.

[0069] Global average pooling compresses the interfering magnetic field features on each channel of the convolutional layer into a scalar, transforming them into... A matrix of size, where This represents the number of channels. Then, this will be... A matrix of size is input into a one-dimensional convolutional layer, which has 1 channel and a kernel size of [size missing]. With the number of channels Related:

[0070]

[0071] For the closest When the odd number of channels is used in the ECA module to calculate the weights of each channel using the SoftMax function, a weight of 0 to 1 is assigned to each interference magnetic field characteristic input to the SoftMax function. The SoftMax function is:

[0072]

[0073] in, Let be the interference magnetic field feature of the i-th channel after convolution and global average pooling, where n is the number of channels. The SoftMax function exponentially transforms the interference magnetic field feature of the i-th channel. Divide by the sum of the exponentially induced interference magnetic field characteristics of all channels, and... The value is mapped to (0,1), which is the weight of the i-th channel. Finally, the attention weight is multiplied by the interference magnetic field characteristics of different channels to obtain the output of the attention layer.

[0074] The diurnal variation network consists of four fully connected layers. The first three fully connected layers have 20 neurons and use the tanh activation function. The last fully connected layer has only one neuron and uses a linear activation function. It takes a normalized time variable as input to fit the diurnal variation of the magnetic field and outputs the fitted diurnal variation value of the geomagnetic field.

[0075] The fourth step is to train the aeromagnetic compensation architecture using the dataset, and retain the interference network as the aeromagnetic compensation model after training.

[0076] During training, training groups are extracted from the training set, and the original variable groups in the groups are input into the aeromagnetic compensation architecture. Specifically, the normalized position variable is input into the geomagnetic network to obtain the fitted geomagnetic field value, the normalized direction cosine variable is input into the interference network to obtain the fitted interference magnetic field value, and the normalized time variable is input into the diurnal variation network to obtain the fitted geomagnetic diurnal variation value. The three are summed to obtain the predicted total field value of the entire model. The predicted total field value is compared with the total magnetic field value in the training group to calculate the loss. The model is updated based on the calculated loss until the entire aeromagnetic compensation architecture converges.

[0077] The validation set data is input into the converged aeromagnetic compensation architecture for verification to ensure the feasibility of the aeromagnetic compensation architecture. Finally, the aeromagnetic compensation architecture that meets the accuracy requirements is output, and the interference network in it is extracted as a separate aeromagnetic compensation model.

[0078] The fifth step involves normalizing the flight data to be processed and then inputting it into the aeromagnetic compensation model to achieve aeromagnetic compensation. Specifically:

[0079] The flight data obtained during the test flight is normalized and then input into the flight compensation network to obtain the predicted magnetic interference value. The predicted magnetic interference value is then removed from the total field magnetic force value measured by the optical pump magnetometer to complete the aeromagnetic compensation operation.

[0080] During test flights, measured data is directly input into the interference network to obtain predicted magnetic interference values. Magnetic interference is then removed from the subsequent total field data, thus completing the compensation operation.

[0081] After compensation is completed, the compensation result is quantitatively evaluated using the improvement ratio (IR). IR is the ratio of the standard deviation of the uncompensated magnetic field after bandpass filtering to the standard deviation of the compensated magnetic field after bandpass filtering.

[0082] To evaluate the effectiveness of the proposed method, calibration and test flights were conducted. Both flights followed a rectangular path with successively orthogonal headings, at an altitude of approximately 2500m with an altitude variation of approximately 150m. During the flights, small-amplitude maneuvers of pitch (±5°), roll (±10°), and yaw (±5°) were repeatedly performed 4-6 times in each direction.

[0083] To compare the effectiveness of the proposed method, it is compared with existing BPNN, GRNN, ResNet, and market-ready product (RMS) compensation results. The input of the magnetic interference network in the BP method is the same as that in this paper, and the compensation effect comparison is shown in Table 1.

[0084] Table 1 Comparison of Compensation Effects

[0085]

[0086] In calibration flight experiments, the proposed method achieved improvements of 113%, 81%, and 164% compared to BPNN, GRNN, and ResNet, respectively, and also showed an 11% improvement compared to the RMS aeromagnetic compensator. In validation flight experiments, the proposed method achieved improvements of 107%, 813%, and 419% compared to BPNN, GRNN, and ResNet, respectively. This demonstrates that the proposed method can effectively improve the accuracy of neural network-based aeromagnetic compensation methods. Since the network trained in the calibration flight experiments was used for validation flights, the compensation effect in the validation flights decreased. The decrease rates for BPNN, GRNN, ResNet, and the proposed method were 15%, 84%, 58%, and 18%, respectively. Although the generalization ability of the proposed method is not optimal, its overall performance in terms of accuracy and generalization ability is the best.

[0087] Please see Figure 5 , Figure 5 This is a schematic diagram of the hardware device in operation according to an embodiment of the present invention. The hardware device specifically includes: a computer device 401, a processor 402, and a storage medium 403.

[0088] A computer device 401: The computer device 401 implements the neural network-based aeromagnetic compensation method.

[0089] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the neural network-based aeromagnetic compensation method.

[0090] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the neural network-based aeromagnetic compensation method.

[0091] The beneficial effects of this invention are as follows: By refining the aeromagnetic compensation architecture based on neural networks and using a 1D CNN-Attention model in the interference network, this invention improves the generalization ability of the convolutional network by adding an ECA attention mechanism between convolutional layers, thereby improving the compensation effect of the verification flight experiment. It can also have a good aeromagnetic interference compensation effect in complex magnetic measurement environments, thus improving the quality of airborne geophysical data.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A neural network-based aeromagnetic compensation method, characterized in that, The specific steps include: S1. Calibration flight, acquire calibration flight data; S2. Select time points, process and calibrate flight data to construct input groups, and construct and divide datasets based on input groups. The input group includes the total magnetic field value and the normalized results of time, direction cosine, and spatial variables obtained from the processed calibration flight data; S3. Establish an aeromagnetic compensation architecture based on neural networks, specifically including geomagnetic networks, interference networks, and diurnal variation networks; S4. Use the dataset to train the aeromagnetic compensation architecture, and retain the interference network as the aeromagnetic compensation model after training. S5. Normalize the flight data to be processed and input it into the aeromagnetic compensation model to achieve aeromagnetic compensation.

2. The aeromagnetic compensation method based on a neural network as described in claim 1, characterized in that, The calibration flight is the flight of the aircraft at a preset flight altitude and the range of flight altitude changes, during which small-amplitude maneuvers such as pitch, roll, and yaw are repeatedly performed. The calibration flight data includes the aircraft's attitude, position, and time information, as well as measurements from the three-axis fluxgate magnetometer and the optical pump magnetometer.

3. The aeromagnetic compensation method based on a neural network as described in claim 2, characterized in that, Step S2 is as follows: S21. Select time points during the calibration flight process according to preset time intervals; S22. Process the calibration flight data corresponding to the time point and construct the original variable set corresponding to the time point; S23. Obtain the optical pump magnetometer measurement values ​​at time points as the total field magnetic force value; S24. Normalize all data in the original variable groups according to the variable type, and then combine the total magnetic field value corresponding to the time point as the training group. Construct a dataset with the training group and divide it into training set and validation set.

4. The aeromagnetic compensation method based on a neural network as described in claim 3, characterized in that, Step S22 is as follows: S221. Combine the selected time point with the square of the calculated time point as the time variable; S222. Combine the aircraft attitude information corresponding to the selected time point with the measurement value of the three-axis fluxgate magnetometer to calculate the direction cosine, and calculate and obtain the quadratic term and the differential term of the direction cosine, and combine them as the direction cosine variable. S223. Select the longitude, latitude, and altitude from the aircraft's location information corresponding to the selected time point, and combine the calculated squared value of longitude, squared value of latitude, and the product of longitude and latitude as the location variable; S224. Combine the time variable, direction cosine variable, and position variable at the selected time point as a set of original variables.

5. The aeromagnetic compensation method based on a neural network as described in claim 1, characterized in that, The geomagnetic network is a one-dimensional convolutional neural network, consisting of convolutional layers, activation layers, pooling layers, flattening layers, and fully connected layers. The input is a normalized position variable, and the output is a fitted geomagnetic field value.

6. The aeromagnetic compensation method based on a neural network as described in claim 1, characterized in that, The interference network consists of a one-dimensional convolutional neural network and an ECA module. The input is a normalized direction cosine variable, and the output is the fitted interference magnetic field value.

7. The aeromagnetic compensation method based on a neural network as described in claim 1, characterized in that, The diurnal variation network consists of four fully connected layers. The first three fully connected layers have 20 neurons and use the tanh activation function. The last fully connected layer has only one neuron and uses the linear activation function. The input is a normalized time variable, and the output is the fitted geomagnetic diurnal variation value.

8. The aeromagnetic compensation method based on a neural network as described in claim 1, characterized in that, Step S5 is as follows: The flight data obtained during the test flight is normalized and then input into the flight compensation network to obtain the predicted magnetic interference value. The predicted magnetic interference value is then removed from the total field magnetic force value measured by the optical pump magnetometer to complete the aeromagnetic compensation operation.

9. A storage medium, characterized in that: The storage medium stores instructions and data to implement the aeromagnetic compensation method based on a neural network as described in any one of claims 1 to 8.

10. A computer device, characterized in that: include: A processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the aeromagnetic compensation method based on a neural network as described in any one of claims 1 to 8.

Images