Self-adaptive weighted double-current aeromagnetic compensation method based on spring model

Abstract

The invention belongs to the field of geomagnetic navigation, and particularly relates to a self-adaptive weighted double-current aeromagnetic compensation method based on a spring model. The method comprises the following steps: collecting geomagnetic information and current information from an aircraft sensor, and generating a corresponding data set; a Tools-Lawson model is used for calibration, and an interference field of the aircraft is removed; a self-adaptive weighted double-current model based on a spring model is utilized, geomagnetic information and current information are fused, comprehensive semantic information is captured through a time attention mechanism, and compensation of an interference field of an airborne electronic system is achieved. And training is carried out in combination with the Tools-Lawson model and the SAWD model, and the trained model can realize accurate compensation of aeromagnetic data. According to the method, the problem that weight adjustment is rigid in an existing aeromagnetic compensation method is effectively solved, and the model can dynamically adapt to different scenes.

Description

Technical Field

[0001] The invention belongs to the field of geomagnetic navigation, and in particular relates to an adaptive weighted dual-flow aeromagnetic compensation method based on a spring model. Background Art

[0002] Aircraft geomagnetic navigation refers to the use of geomagnetic data to achieve precise positioning and route planning. Currently, mainstream aircraft navigation systems mainly include inertial navigation systems (INS) and global navigation satellite systems (GNSS), such as GPS, BDS, and GLONASS. However, GNSS relies heavily on satellite signals, and its performance degrades significantly in environments with weak or blocked signals. Although INS is of great value, its cumulative error increases over time, and high-precision INS requires expensive inertial sensors, which greatly increases the cost. These limitations have led to growing interest in aircraft geomagnetic navigation, as geomagnetic navigation provides a more cost-effective alternative. Unlike traditional methods, geomagnetic navigation does not require expensive infrastructure and can operate in environments without satellite signals, such as during military operations. This research has important theoretical and practical significance for improving the capabilities of navigation systems in complex environments.

[0003] Despite the potential of geomagnetic navigation, its low positioning accuracy has prevented widespread adoption. This is primarily due to errors in real-time geomagnetic readings and poor navigation algorithm performance. This challenge stems from the core principle of aircraft geomagnetic navigation: comparing magnetic field data collected by an onboard magnetometer with a geomagnetic reference map to enable autonomous navigation or to assist other navigation systems (such as INS).

[0004] In practical applications, reliable aeromagnetic compensation methods face several significant challenges, including low accuracy, poor real-time performance, and limited versatility. First, real-time geomagnetic data suffers from diverse sources of error. In addition to systematic errors in magnetometer readings, these errors also come from the aircraft's magnetic field, electromagnetic interference within the cockpit, and interference from the geomagnetic field during low-altitude flight. Because aircraft are typically far above the ground during flight, geomagnetic interference is minimal, so interference from the aircraft's magnetic field and cockpit electromagnetic field is primarily considered. However, the intensity of these interference fields varies over time, making simple mathematical methods difficult to accurately model. Second, during high-speed flight, slow geomagnetic navigation algorithms can cause a lag in the position information provided to the user, resulting in significant deviations between the aircraft's actual and reported positions, and consequently, navigation decisions that are inconsistent with the actual state. Furthermore, some methods only effectively eliminate geomagnetic interference on specific routes but perform poorly on other routes. This lack of robustness and versatility limits their applicability to a wider range of flight data. Addressing these issues and achieving more accurate and reliable magnetic field compensation is a key technical challenge in the field. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem that the existing aeromagnetic compensation method cannot remove the electromagnetic interference generated by the aircraft's onboard electronic system, and to propose an adaptive weighted dual-flow aeromagnetic compensation method based on a spring model.

[0006] In order to achieve the above object, the technical solution provided by the present invention is:

[0007] An adaptive weighted dual-flow aeromagnetic compensation method based on a spring model, comprising:

[0008] Acquire calibration geomagnetic data of the calibration aircraft, and acquire real-time geomagnetic data and real-time current data of the navigation aircraft, wherein the real-time geomagnetic data includes real-time scalar geomagnetic data and real-time vector geomagnetic data;

[0009] Construct an aeromagnetic compensation model, obtain the initial aeromagnetic compensation model coefficients based on the calibrated geomagnetic data, and use the aeromagnetic compensation model to obtain the aircraft interference field based on the real-time vector geomagnetic data;

[0010] An adaptive weighted dual-current model based on a spring model is constructed to obtain the current interference field;

[0011] Jointly train the aeromagnetic compensation model and the adaptive weighted dual-flow model based on the spring model, and synchronously update the aeromagnetic compensation model coefficients, the parameters of the adaptive weighted dual-flow model based on the spring model, and the geomagnetic weights of the real-time scalar geomagnetic data to obtain a trained joint model;

[0012] The trained joint model is used to remove the aircraft interference field and the current interference field to obtain accurate geomagnetic data.

[0013] Furthermore, the aeromagnetic compensation model is constructed, initial aeromagnetic compensation model coefficients are obtained based on the calibrated geomagnetic data, and the aircraft interference field is obtained based on the real-time vector geomagnetic data using the aeromagnetic compensation model, including:

[0014] The calibration geomagnetic data includes calibration scalar geomagnetic data and calibration vector geomagnetic data;

[0015] The aircraft interference field is modeled and expressed as follows:

[0016]

[0017] in, represents the calibrated scalar geomagnetic data, represents the real scalar geomagnetic data after removing the aircraft interference field, represents the coefficient matrix associated with vector geomagnetic data, represents the aeromagnetic compensation model coefficient;

[0018] The real geomagnetic data is filtered by a bandpass filter and expressed as:

[0019]

[0020] in, represents a bandpass filter;

[0021] The coefficient matrix is ​​calculated based on the calibration vector geomagnetic data, and the initial aeromagnetic compensation model coefficients are solved by linear regression;

[0022] The real-time vector geomagnetic data is input into the aeromagnetic compensation model, which calculates the coefficient matrix based on the real-time vector geomagnetic data and outputs the aircraft interference field.

[0023] Furthermore, the adaptive weighted dual-flow model based on the spring model includes:

[0024] Two sets of parallel fully connected layers and batch normalization layers are used to input real-time scalar geomagnetic data and real-time current data respectively;

[0025] A feature fusion layer is used to fuse the real-time scalar geomagnetic data and the real-time current data according to the geomagnetic weight to obtain fused data;

[0026] Two dilated attention layers for feature extraction of fused data;

[0027] A fully connected layer to output the current interference field.

[0028] Furthermore, the dilated attention layer includes: a dilated convolutional neural network with a Swish activation function, a dilated convolutional neural network with a Sigmoid activation function, and a one-dimensional convolutional neural network with a Swish activation function;

[0029] The fused data is input into the dilated convolutional neural network with Swish activation function to obtain global features;

[0030] Input the fused data into the dilated convolutional neural network with Sigmoid activation function to obtain the gating weights;

[0031] Combine the gate weights with the global features through a dot product operation and combine them with the fusion data;

[0032] The combined data is input into a one-dimensional convolutional neural network with a Swish activation function to obtain the output of the dilated attention layer.

[0033] Furthermore, the joint training includes:

[0034] The real-time scalar geomagnetic data and the real-time current data are input into the adaptive weighted dual-current model based on the spring model, and the adaptive weighted dual-current model based on the spring model outputs the current interference field according to the geomagnetic weight adjusted in the previous iteration;

[0035] Update the coefficients of the aeromagnetic compensation model and the parameters of the adaptive weighted dual-flow model based on the spring model according to the total loss function;

[0036] The geomagnetic weight of the current iteration is adjusted using an adaptive weighting strategy based on a spring model;

[0037] The joint training is completed after I iterations.

[0038] Furthermore, the adaptive weighting strategy based on the spring model is used to adjust the geomagnetic weight of the current iteration, including:

[0039] The stiffness coefficient of the initial spring of the spring model is , the initial reading on the right end of the spring is 0.5, and the range of force it can withstand is , when the applied force exceeds When the spring is replaced with the second state spring, the stiffness coefficient of the second state spring is The range of the force that can be withstood is , and so on, when the applied force exceeds When the force that the state spring can withstand is within the range, the spring is replaced with the State Spring, The stiffness coefficient of the state spring is The range of the force that can be withstood is ;

[0040] The reading at the right end of the spring is expressed as:

[0041]

[0042] in, Indicates the reading at the right end of the spring, represents the force applied to the spring, Indicates the The minimum force that the state spring can withstand, Indicates the The maximum force that the state spring can withstand, Indicates the The minimum force that the state spring can withstand, Indicates the State: The maximum force that the spring can withstand;

[0043] The right end reading of the spring in each training iteration is used as the geomagnetic weight;

[0044] Calculate the loss function when only real-time scalar geomagnetic data is input into the adaptive weighted two-flow model based on the spring model and the loss function when only real-time current data is input into the adaptive weighted dual-flow model based on the spring model ;

[0045] According to the loss function and loss function Adjust the force applied to the spring to obtain the geomagnetic weight in the current iteration.

[0046] Furthermore, the loss function and loss function Adjust the force applied to the spring to obtain the geomagnetic weight in the current iteration:

[0047] The first stage is before the spring reaches its elastic limit for the first time. , the following rules apply:

[0048] if , then by reducing The value of reduce;

[0049] if , then keep constant;

[0050] if , then increase The value of improve;

[0051] After the spring has first reached its elastic limit, it enters the second stage, where the following rules apply:

[0052] if , then increase The value of improve;

[0053] if , then keep constant;

[0054] if , then by reducing The value of reduce;

[0055] in, Indicates the geomagnetic weight, the value is equal to the reading at the right end of the spring.

[0056] Furthermore, the loss function when only real-time scalar geomagnetic data is input into the adaptive weighted dual-flow model based on the spring model is: and the loss function when only real-time current data is input into the adaptive weighted dual-flow model based on the spring model The calculation formula is as follows:

[0057]

[0058]

[0059] in, represents the loss function when the input is only real-time scalar geomagnetic data, represents the loss function when the input is only real-time current data, Indicates the number of timestamps, represents real-time scalar geomagnetic data, represents the aircraft interference field output by the aeromagnetic compensation model, represents the predicted value of the current interference field output by the adaptive weighted dual-current model based on the spring model when the input is only real-time scalar geomagnetic data, represents real-time scalar geomagnetic data, represents the current interference field prediction value output by the adaptive weighted dual-current model based on the spring model when the input is only real-time current data, Indicates real-time current data, Represents real scalar geomagnetic data.

[0060] Furthermore, the total loss function is expressed as follows:

[0061]

[0062] in, represents the total loss function, Indicates the number of timestamps, represents real-time scalar geomagnetic data, represents the aircraft interference field output by the aeromagnetic compensation model, represents the current interference field prediction value output by the adaptive weighted dual-current model based on the spring model in the last training iteration, represents the input data of the adaptive weighted two-stream model based on the spring model in the last training iteration, Represents real scalar geomagnetic data.

[0063] Furthermore, the geomagnetic weights of each iteration are averaged, and the average value is used as the geomagnetic weight of the trained joint model. The calculation process is expressed as follows:

[0064]

[0065] in, represents the geomagnetic weight of the trained joint model, Representative The geomagnetic weight in the iteration, I is the total number of iterations in the training phase.

[0066] Compared with the existing technology, the present invention has the following significant advantages: the present invention effectively solves the problem of rigid weight adjustment in existing aeromagnetic compensation methods through an adaptive weighting strategy based on a spring model, enabling the model to dynamically adapt in different scenarios; at the same time, through the combination of dilated convolution and multi-layer temporal attention mechanism, the real-time and robustness of aeromagnetic compensation are significantly improved, and the accuracy of aeromagnetic compensation can be significantly improved under various flight conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is a flow chart of an adaptive weighted dual-flow aeromagnetic compensation method based on a spring model according to the present invention;

[0068] Figure 2 This is a diagram showing the architecture of the aeromagnetic compensation model of the present invention;

[0069] Figure 3 This is an architecture diagram of the spring-based adaptive weighted dual-flow model of the present invention;

[0070] Figure 4 This is the architectural diagram of the expanded attention layer of the present invention. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0072] The present invention provides an adaptive weighted dual-flow aeromagnetic compensation method based on a spring model, which is used to compensate scalar geomagnetic data collected on an aircraft to remove the magnetic interference field of the aircraft itself and the magnetic interference field generated by the power system on the aircraft.

[0073] like Figure 1 As shown, this embodiment includes the following steps:

[0074] 1) Collect geomagnetic information and current information from aircraft sensors to generate corresponding data sets.

[0075] Aircraft sensors include scalar magnetometers, vector magnetometers, and current sensors installed on the navigation aircraft, as well as scalar magnetometers and vector magnetometers installed on the calibration aircraft. Current sensors monitor currents that may interfere with the geomagnetic field. Sources of current include the aircraft's radio equipment, flashlights, system output power supplies, and air conditioning fans. Unlike data collected in real time on the navigation aircraft, data on the calibration route must be collected in advance. The calibration aircraft's flight path should avoid sources of electromagnetic interference, such as power lines, radar stations, and radio towers. The aircraft should fly at a higher altitude to minimize magnetic interference from ground-based infrastructure. Calibration flights should be conducted in an environment with stable geomagnetic conditions, avoiding geomagnetic storms or other disturbances that could affect geomagnetic stability. The flight attitude should remain stable during flight to avoid additional errors introduced by extreme maneuvers. Furthermore, due to differences in the permanent field, induced field, and eddy current field characteristics between aircraft, the calibration aircraft should have the same or highly similar magnetic field characteristics as the navigation aircraft.

[0076] The scalar magnetometer on the calibration aircraft measures calibrated scalar geomagnetic data, the vector magnetometer on the calibration aircraft measures calibrated vector geomagnetic data, the scalar magnetometer on the navigation aircraft measures real-time scalar geomagnetic data, the vector magnetometer on the navigation aircraft measures real-time vector geomagnetic data, and the current sensor on the navigation aircraft measures real-time current data.

[0077] 2) Use the aeromagnetic compensation model to calibrate and remove the aircraft interference field.

[0078] The method for removing aircraft interference fields is based on the traditional aeromagnetic compensation model: the Tolles-Lawson model (TL model for short). The Tolles-Lawson model uses relevant physical knowledge to model aircraft interference fields. Its core formula is:

[0079] ;

[0080] in, represents the calibrated scalar geomagnetic data, represents the real scalar geomagnetic data after removing the aircraft interference field, represents the coefficient matrix (TL matrix) associated with vector geomagnetic data, Represents the aeromagnetic compensation model coefficient (TL coefficient), which is the vector that needs to be solved in the TL model.

[0081] The interference field and TL coefficient Since the values ​​are unknown, the formula cannot be solved. Therefore, it is necessary to filter out the real scalar geomagnetic data through a bandpass filter (bpf), which is why the calibration route is required in step 1). With this setting, the bandpass filter can remove almost all of the Earth's field. , while retaining most of the aircraft interference field, namely:

[0082] ;

[0083] The TL coefficient can be obtained by linear regression. Solve to get the initial TL coefficient . In getting The aircraft interference field can then be obtained by combining it with the real-time vector geomagnetic data.

[0084] In this embodiment, if Figure 2 As shown, the TL matrix is ​​first calculated using the calibration vector geomagnetic data collected in advance. , The direction cosine row vector Composed in time series:

[0085] ;

[0086] in, represents the length of the time series, and is a row vector with 18 entries:

[0087] ;

[0088] in, 、 and Is vector geomagnetic data The direction cosines of represents the derivative with respect to time, represents the unit vector, Represents the magnitude of the vector.

[0089] Then, the calibrated scalar geomagnetic data are passed through a bandpass filter to remove the Earth field component in the calibrated scalar geomagnetic data, and the TL coefficient is calculated by linear regression. After obtaining the TL coefficient, the TL matrix obtained by combining the real-time vector geomagnetic data is The aircraft interference field can be obtained.

[0090] 3) Construct an adaptive weighted dual-flow (SAWD) model based on the spring model. Using the SAWD model, the geomagnetic information and current information are integrated, and the comprehensive semantic information is captured through the temporal attention mechanism to compensate for the interference field (current interference field) of the airborne electronic system. Figure 3 As shown, specifically including:

[0091] 3.1) The SAWD model includes two parallel fully connected layers and batch normalization layers, which are used to input real-time scalar geomagnetic data and real-time current data respectively; a feature fusion layer, which is used to fuse the real-time scalar geomagnetic data and real-time current data according to the geomagnetic weight to obtain fused data; two dilated attention layers, which are used to extract features from the fused data; and a fully connected layer, which is used to output the current interference field.

[0092] 3.2) Take the real-time scalar geomagnetic data and real-time current data collected by the navigation aircraft as input, and use The data is then processed using a batch normalization layer to ensure that the generated outputs have the same size, thus achieving data alignment before merging.

[0093] 3.3) An adaptive weighting strategy based on a spring model is used to dynamically adjust the geomagnetic weights of real-time scalar geomagnetic data and real-time current data during the training process.

[0094] The adaptive weighting strategy based on the spring model is to simulate the geomagnetic weight adjustment strategy into a spring model. The characteristics of the proposed spring model are: the initial spring Ω1, and its stiffness coefficient is At the initial moment, the spring is not stressed (i.e. =0), the right end of the spring reads 0.5. The force range that the spring can withstand is According to Hooke's law, the elastic limit range of this stage is When the applied force exceeds this range, the spring is immediately replaced with a stiffer spring Ω2, whose stiffness coefficient is increased by an increment , forming a new stiffness coefficient Therefore, the range of forces that the spring can withstand becomes If the force range returns to , then it is replaced with Ω1 again. If the applied force exceeds the support range of Ω1 and Ω2, the spring is replaced with Ω3, whose stiffness is increased to , the range of forces that can be tolerated is extended to This process continues with each subsequent spring Ω In the force range of is used, and its stiffness coefficient is These variables satisfy the following relationship:

[0095] ;

[0096] ;

[0097] ;

[0098] in, Indicates the The stiffness coefficient of the state spring, Indicates the The minimum force that the state spring can withstand, Indicates the The maximum force that the state spring can withstand, is the increment of the reciprocal of the stiffness coefficient, The critical range expansion increment is The value range of is a positive integer. Based on the above description, we can further deduce the quantitative relationship between the reading at the right end of the spring and the applied force, which can be expressed as:

[0099] ;

[0100] in, is a positive integer, Indicates the reading at the right end of the spring, represents the force applied to the spring, Indicates the The minimum force that the state spring can withstand, Indicates the The maximum force that the state spring can withstand.

[0101] In each training iteration the right end reading of the spring is assigned a geomagnetic weight, denoted as , the corresponding real-time current data weight is expressed as (1- ). In the initial iteration, is set to 0.5. In subsequent iterations, the loss values ​​in the two cases are calculated separately. and :One is to use only real-time scalar geomagnetic field data, and the other is to use only real-time current data. The loss function is calculated as:

[0102] ;

[0103] ;

[0104] in, represents the loss function when the input is only real-time scalar geomagnetic data, represents the loss function when the input is only real-time current data, Indicates the number of timestamps, represents real-time scalar geomagnetic data, represents the aircraft interference field output by the aeromagnetic compensation model, represents the predicted value of the current interference field output by the SAWD model when the input is only real-time scalar geomagnetic data, represents real-time scalar geomagnetic data, It represents the predicted value of the current interference field output by the SAWD model when the input is only real-time current data. Indicates real-time current data, Represents real scalar geomagnetic data.

[0105] Adjustment of spring parameters is based on and This process is divided into two stages:

[0106] In the first stage, before the spring first reaches its elastic limit, the following rules apply:

[0107] if , then by reducing The value of (reduced by ),make reduce;

[0108] if , then keep constant;

[0109] if , then increase The value of (the increase is ),make improve.

[0110] After the spring has first reached its elastic limit, it enters the second phase, where the rules are reversed:

[0111] if , then increase The value of (the increase is ),make improve;

[0112] if , then keep constant;

[0113] if , then by reducing The value (reduction is Δ ),make reduce.

[0114] In the adaptive weighted strategy based on the spring model, and Respectively represent the influence of geomagnetic information and current information on the compensation value. , indicating that geomagnetic information is more accurate in estimating the compensation value, reflecting the importance of geomagnetic information. Therefore, before the spring reaches its elastic limit for the first time, increasing the geomagnetic weight This helps the SAWD model better capture the overall pattern of geomagnetic information. When the spring reaches its elastic limit, the SAWD model has fully learned the geomagnetic information and has converged to a certain extent based on this data, but the learning of current information is still insufficient. At this time, reducing This helps fine-tune the SAWD model to better handle more complex current disturbances. Secondly, when the model reaches its elastic limit for the second time, using a spring with a smaller stiffness coefficient can help the SAWD model quickly adapt to the new training phase and achieve convergence.

[0115] 3.4) Based on the adjusted geomagnetic weights, the two objects processed by the batch normalization layer in step 3.1) are fused to generate fused data. Two dilated attention layers are used to further extract features from the fused data, and a fully connected layer is used to generate information about the current interference field generated by the airborne electronic system.

[0116] Expanding the attention layer Figure 4 As shown, specifically including:

[0117] The fused information is processed by two dilated convolutional neural networks. Among them, the dilated convolutional neural network with the Swish activation function is used to extract the global features of the fused data, while the dilated convolutional neural network with the Sigmoid activation function is used as a gating mechanism to map the data to the range of [0,1] as the gating weight; the above gating weights are then combined with the global features through a dot product operation to achieve a deep fusion of geomagnetic information and current information, and combine the input data to retain more original information; the fused information is further processed by a one-dimensional convolutional neural network with the Swish activation function to enhance the learning ability of the model.

[0118] In the SAWD model, dilated convolutions are chosen for their larger receptive field and lower computational cost. The dilation rates of different dilated attention layers vary, and they decrease as data is passed between layers. This design aims to minimize information loss in the SAWD model.

[0119] 4) Combine the TL model and the SAWD model for joint training. The trained joint model can achieve accurate compensation of aeromagnetic data. Specifically, it includes:

[0120] Based on the fusion data obtained from the geomagnetic weight obtained from the previous iterative training, the aeromagnetic compensation model coefficients are trained jointly. and SAWD model parameters, and dynamically adjust the geomagnetic weights based on the spring model to finally obtain the trained joint model.

[0121] Joint training is performed based on the following loss function:

[0122]

[0123] in, represents the total loss function, represents the current interference field prediction value output by the SAWD model in the last training iteration, Represents the input data of the SAWD model in the last training iteration.

[0124] In one iteration, the parameters of the SAWD model and the TL coefficients are first adjusted based on the fusion data of the previous iteration. Then use real-time scalar geomagnetic data and real-time current data as input, according to and The total number of iterations is 1.

[0125] It should be noted that: Although increasing the spring stiffness during training can effectively prevent is close to 0 or 1, but when the number of iterations is large, It is still possible that θ will tend towards these extreme values, or even reach 0 or 1 at the end of training. The parameters are relatively random. Therefore, in the test phase, the last iteration of the training phase is not used directly. value, but rather by performing The values ​​are averaged to calculate the test model This approach ensures that the test model The parameters are more stable and more fully represent the entire training process.

[0126] Testing phase The value can be calculated using the following formula:

[0127] ;

[0128] in, Indicates the testing phase value, Representative In the iteration value, and I is the total number of iterations in the training phase.

[0129] Real-time geomagnetic data and real-time current data are input into the trained joint model. The TL model is used to eliminate the aircraft interference field generated by the aircraft itself in the real-time geomagnetic data, and the SAWD model is used to remove the current interference field generated by the aircraft power system. Finally, accurate geomagnetic data with compensation is output for geomagnetic navigation.

[0130] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. An adaptive weighted dual-flow aeromagnetic compensation method based on a spring model, characterized in that: The adaptive weighted dual-flow aeromagnetic compensation method based on the spring model includes: Acquire calibration geomagnetic data of the calibration aircraft, and acquire real-time geomagnetic data and real-time current data of the navigation aircraft, wherein the real-time geomagnetic data includes real-time scalar geomagnetic data and real-time vector geomagnetic data; Construct an aeromagnetic compensation model, obtain the initial aeromagnetic compensation model coefficients based on the calibrated geomagnetic data, and use the aeromagnetic compensation model to obtain the aircraft interference field based on the real-time vector geomagnetic data; An adaptive weighted dual-current model based on a spring model is constructed to obtain the current interference field; Jointly train the aeromagnetic compensation model and the adaptive weighted dual-flow model based on the spring model, and synchronously update the aeromagnetic compensation model coefficients, the parameters of the adaptive weighted dual-flow model based on the spring model, and the geomagnetic weights of the real-time scalar geomagnetic data to obtain a trained joint model; The trained joint model is used to remove the aircraft interference field and the current interference field to obtain accurate geomagnetic data.

2. The adaptive weighted dual-flow aeromagnetic compensation method based on the spring model according to claim 1 is characterized in that: The aeromagnetic compensation model is constructed, initial aeromagnetic compensation model coefficients are obtained based on calibrated geomagnetic data, and aircraft interference fields are obtained based on real-time vector geomagnetic data using the aeromagnetic compensation model, including: The calibration geomagnetic data includes calibration scalar geomagnetic data and calibration vector geomagnetic data; The aircraft interference field is modeled and expressed as follows: in, represents the calibrated scalar geomagnetic data, represents the real scalar geomagnetic data after removing the aircraft interference field, represents the coefficient matrix associated with vector geomagnetic data, represents the aeromagnetic compensation model coefficient; The real geomagnetic data is filtered by a bandpass filter and expressed as: in, represents a bandpass filter; The coefficient matrix is ​​calculated based on the calibration vector geomagnetic data, and the initial aeromagnetic compensation model coefficients are solved by linear regression; The real-time vector geomagnetic data is input into the aeromagnetic compensation model, which calculates the coefficient matrix based on the real-time vector geomagnetic data and outputs the aircraft interference field.

3. The adaptive weighted dual-flow aeromagnetic compensation method based on the spring model according to claim 1 is characterized in that: The adaptive weighted dual-flow model based on the spring model includes: Two sets of parallel fully connected layers and batch normalization layers are used to input real-time scalar geomagnetic data and real-time current data respectively; A feature fusion layer is used to fuse the real-time scalar geomagnetic data and the real-time current data according to the geomagnetic weight to obtain fused data; Two dilated attention layers for feature extraction of fused data; A fully connected layer to output the current interference field.

4. The adaptive weighted dual-flow aeromagnetic compensation method based on the spring model according to claim 3 is characterized in that: The dilated attention layer includes: a dilated convolutional neural network with a Swish activation function, a dilated convolutional neural network with a Sigmoid activation function, and a one-dimensional convolutional neural network with a Swish activation function; The fused data is input into the dilated convolutional neural network with Swish activation function to obtain global features; Input the fused data into the dilated convolutional neural network with Sigmoid activation function to obtain the gating weights; Combine the gate weights with the global features through a dot product operation and combine them with the fusion data; The combined data is input into a one-dimensional convolutional neural network with a Swish activation function to obtain the output of the dilated attention layer.

5. The adaptive weighted dual-flow aeromagnetic compensation method based on a spring model according to claim 1, characterized in that: The joint training includes: The real-time scalar geomagnetic data and the real-time current data are input into the adaptive weighted dual-current model based on the spring model, and the adaptive weighted dual-current model based on the spring model outputs the current interference field according to the geomagnetic weight adjusted in the previous iteration; Update the coefficients of the aeromagnetic compensation model and the parameters of the adaptive weighted dual-flow model based on the spring model according to the total loss function; The geomagnetic weight of the current iteration is adjusted using an adaptive weighting strategy based on a spring model; The joint training is completed after I iterations.

6. The adaptive weighted dual-flow aeromagnetic compensation method based on a spring model according to claim 5, characterized in that: The method of adjusting the geomagnetic weight of the current iteration using the adaptive weighting strategy based on the spring model includes: The stiffness coefficient of the initial spring of the spring model is , the initial reading on the right end of the spring is 0.5, and the range of force it can withstand is , when the applied force exceeds When the spring is replaced with the second state spring, the stiffness coefficient of the second state spring is The range of the force that can be withstood is , and so on, when the applied force exceeds When the force that the state spring can withstand is within the range, the spring is replaced with the State Spring, The stiffness coefficient of the state spring is The range of the force that can be withstood is ; The reading at the right end of the spring is expressed as: in, Indicates the reading at the right end of the spring, represents the force applied to the spring, Indicates the The minimum force that the state spring can withstand, Indicates the The maximum force that the state spring can withstand, Indicates the The minimum force that the state spring can withstand, Indicates the State: The maximum force that the spring can withstand; The right end reading of the spring in each training iteration is used as the geomagnetic weight; Calculate the loss function when only real-time scalar geomagnetic data is input into the adaptive weighted two-flow model based on the spring model and the loss function when only real-time current data is input into the adaptive weighted dual-flow model based on the spring model ; According to the loss function and loss function Adjust the force applied to the spring to obtain the geomagnetic weight in the current iteration.

7. The adaptive weighted dual-flow aeromagnetic compensation method based on a spring model according to claim 6, characterized in that: According to the loss function and loss function Adjust the force applied to the spring to obtain the geomagnetic weight in the current iteration: The first stage is before the spring reaches its elastic limit for the first time. , the following rules apply: if , then by reducing The value of reduce; if , then keep constant; if , then increase The value of improve; After the spring has first reached its elastic limit, it enters the second stage, where the following rules apply: if , then increase The value of improve; if , then keep constant; if , then by reducing The value of reduce; in, Indicates the geomagnetic weight, the value is equal to the reading at the right end of the spring.

8. The adaptive weighted dual-flow aeromagnetic compensation method based on a spring model according to claim 6, characterized in that: The loss function when only real-time scalar geomagnetic data is input into the adaptive weighted dual-flow model based on the spring model and the loss function when only real-time current data is input into the adaptive weighted dual-flow model based on the spring model The calculation formula is as follows: in, represents the loss function when the input is only real-time scalar geomagnetic data, represents the loss function when the input is only real-time current data, Indicates the number of timestamps, represents real-time scalar geomagnetic data, represents the aircraft interference field output by the aeromagnetic compensation model, represents the predicted value of the current interference field output by the adaptive weighted dual-current model based on the spring model when the input is only real-time scalar geomagnetic data, represents real-time scalar geomagnetic data, represents the current interference field prediction value output by the adaptive weighted dual-current model based on the spring model when the input is only real-time current data, Indicates real-time current data, Represents real scalar geomagnetic data.

9. The adaptive weighted dual-flow aeromagnetic compensation method based on a spring model according to claim 5, characterized in that: The total loss function is expressed as follows: in, represents the total loss function, Indicates the number of timestamps, represents real-time scalar geomagnetic data, represents the aircraft interference field output by the aeromagnetic compensation model, represents the current interference field prediction value output by the adaptive weighted dual-current model based on the spring model in the last training iteration, represents the input data of the adaptive weighted two-stream model based on the spring model in the last training iteration, Represents real scalar geomagnetic data.

10. The adaptive weighted dual-flow aeromagnetic compensation method based on a spring model according to claim 1, characterized in that: The geomagnetic weights of each iteration are averaged and the average value is used as the geomagnetic weight of the trained joint model. The calculation process is expressed as follows: in, represents the geomagnetic weight of the trained joint model, Representative The geomagnetic weight in the iteration, I is the total number of iterations in the training phase.

Images