Unmanned aerial vehicle magnetic gradient data soft compensation method and system based on semi-supervised learning

Abstract

The application discloses a UAV magnetic gradient data soft compensation method and system based on semi-supervised learning, relates to the technical field of aerogeophysical exploration, and comprises the following steps: designing an octagonal calibration flight scheme, a UAV performs a single maneuvering action on each flight line, and response data of an interference magnetic field to flight attitude changes are acquired; measured high-altitude compensation flight data and low-altitude survey line flight data are based on high-altitude compensation flight data and low-altitude survey line flight data in the response data of the interference magnetic field to the flight attitude changes, a simulation dynamic measurement noise data set is established through a forward formula, and the simulation dynamic measurement noise data set is divided into a training set, a verification set and a test set; and an improved octagonal calibration flight scheme is combined with a semi-supervised learning residual error compensation method based on a Transformer to improve soft compensation precision and efficiency.

Description

Technical Field

[0001] This invention relates to the field of airborne geophysical exploration technology, and in particular to a method and system for soft compensation of UAV magnetic gradient data based on semi-supervised learning. Background Technology

[0002] Traditional UAV full-tensor magnetic gradient measurements employ a rectangular flight path calibration scheme, performing pitch, roll, and yaw maneuvers in four orthogonal directions (e.g., east, south, west, and north) to excite interfering magnetic field components.

[0003] Traditional flight schemes often cause yaw maneuvers to lead to the magnetometer entering the signal dead zone, resulting in unstable measurement data. Flight maneuvers are complex and require high skill from pilots. Existing soft compensation methods are based on physical models and cannot suppress all dynamic measurement noise. Traditional supervised learning methods rely on large-scale labeled data, have poor generalization ability, low efficiency in processing high sampling rate data, and the noise has spatiotemporal correlation. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a semi-supervised learning-based soft compensation method for UAV magnetic gradient data to solve the operational complexity and sample dependence problems of traditional schemes, while improving the signal-to-noise ratio of full tensor magnetic gradient data.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, this invention provides a semi-supervised learning-based soft compensation method for UAV magnetic gradient data. The method includes: designing an octagonal calibration flight scheme, whereby the UAV performs a single maneuver on each flight path to acquire response data of the interfering magnetic field to changes in flight attitude; establishing a simulated dynamic measurement noise dataset using forward modeling based on high-altitude compensation flight data and low-altitude survey line flight data from the response data of the interfering magnetic field to changes in flight attitude, and dividing it into training, validation, and test sets; constructing a Transformer-based semi-supervised learning network model to compensate for the interfering magnetic field in the full tensor magnetic gradient data, and calculating the theoretical full tensor magnetic gradient data based on the International Geomagnetic Reference Field Model; designing a multi-constraint loss function; and optimizing the parameters of the Transformer-based semi-supervised learning network model using a stepwise training strategy with the training and validation sets, and verifying the effect based on the test set.

[0008] As a preferred embodiment of the semi-supervised learning-based UAV magnetic gradient data soft compensation method described in this invention, the octagonal calibration flight scheme is designed, and the UAV performs a single maneuver on each flight path. The specific steps are as follows.

[0009] Centered on points with small changes in the geomagnetic gradient, stable areas are selected as measurement sites. Flight paths with different azimuth angles are set up to form an octagonal calibration flight plan. Pitch and roll maneuvers are performed twice to obtain data on the response of the interfering magnetic field to changes in flight attitude.

[0010] As a preferred embodiment of the semi-supervised learning-based UAV magnetic gradient data soft compensation method described in this invention, the specific steps for the high-altitude compensation flight data and low-altitude survey line flight data based on the response data of the interfering magnetic field to changes in flight attitude are as follows:

[0011] The high-altitude compensated flight data in the response data of the interference magnetic field to changes in flight attitude are used as noise-labeled data, and the low-altitude survey line flight data are used as unlabeled data. The mutual conversion between the geographic coordinate system, the geocentric coordinate system and the navigation coordinate system is handled by the coordinate transformation theory formula.

[0012] As a preferred embodiment of the semi-supervised learning-based UAV magnetic gradient data soft compensation method described in this invention, the specific steps for establishing a simulated dynamic measurement noise dataset using forward modeling formulas and dividing it into a training set, a validation set, and a test set are as follows.

[0013] Simulation data is generated in batches by exhaustively combining attitude changes and compensation coefficients using forward modeling formulas. The high-altitude compensated flight data and low-altitude survey line flight data from the response data of the interference magnetic field to flight attitude changes are merged with the simulation data to construct a simulation dynamic measurement noise dataset. The simulation dynamic measurement noise dataset is divided into training set, validation set and test set.

[0014] As a preferred embodiment of the semi-supervised learning-based soft compensation method for UAV magnetic gradient data described in this invention, the semi-supervised learning network model based on Transformer includes an input module, a full tensor magnetic gradient encoder, an error compensation decoder, and an output module.

[0015] By extracting features from multiple error sources through an attention mechanism and compensating for interfering magnetic fields in the full tensor magnetic gradient data, theoretical full tensor magnetic gradient data are calculated based on the international geomagnetic reference field model.

[0016] As a preferred embodiment of the semi-supervised learning-based UAV magnetic gradient data soft compensation method of the present invention, the multi-constraint loss function is composed of a weighted sum of a label data constraint loss function term, a multi-gradient component transformation relationship constraint loss function term, and an overfitting suppression loss function term.

[0017] As a preferred embodiment of the semi-supervised learning-based UAV magnetic gradient data soft compensation method of the present invention, wherein: the label data constraint loss function term is defined based on measuring the difference between the predicted value and the high-altitude compensated flight noise label data; the multi-gradient component transformation relationship constraint loss function term is defined based on the quantitative relationship of the integral transformation between the horizontal magnetic gradient components; and the overfitting suppression loss function term is obtained by balancing the weights of different loss functions in the training of the Transformer semi-supervised learning network model.

[0018] As a preferred embodiment of the semi-supervised learning-based soft compensation method for UAV magnetic gradient data described in this invention, the steps of optimizing the parameters of the Transformer-based semi-supervised learning network model using a stepwise training strategy with training and validation sets, and verifying the effect based on the test set, are as follows:

[0019] The semi-supervised learning network model based on Transformer is trained batch by batch using the training set, and the parameters are continuously adjusted to minimize the multi-constraint loss function.

[0020] The performance of the Transformer-based semi-supervised learning network model is evaluated using a validation set. The multi-constraint loss value on the validation set is calculated, the parameters of the Transformer-based semi-supervised learning network model are optimized through cross-validation, and the optimal parameters of the Transformer-based semi-supervised learning network model with the lowest validation set loss are saved.

[0021] As a preferred embodiment of the semi-supervised learning-based UAV magnetic gradient data soft compensation method described in this invention, the specific steps for verifying the effect based on the test set are as follows:

[0022] The effectiveness of the semi-supervised learning-based aeromagnetic interference calibration method is verified by calculating the multi-constraint loss value and performance index on the test set.

[0023] Secondly, this invention provides a semi-supervised learning-based soft compensation system for UAV magnetic gradient data, comprising a flight scheme design module, a data acquisition module, a network model construction module, a loss function design module, and a training and verification module. The flight scheme design module designs an octagonal calibration flight scheme, whereby the UAV performs a single maneuver on each flight path to acquire response data of the interfering magnetic field to changes in flight attitude. The data acquisition module establishes a simulated dynamic measurement noise dataset using forward modeling formulas based on high-altitude compensation flight data and low-altitude survey line flight data from the response data of the interfering magnetic field to changes in flight attitude, and divides it into training, verification, and test sets. The network model construction module constructs a Transformer-based semi-supervised learning network model to compensate for the interfering magnetic field in the full tensor magnetic gradient data, calculating the theoretical full tensor magnetic gradient data based on the International Geomagnetic Reference Field Model. The loss function design module designs a multi-constraint loss function. The training and verification module optimizes the parameters of the Transformer-based semi-supervised learning network model using a stepwise training strategy with the training and verification sets, and verifies the effect based on the test set.

[0024] The beneficial effects of this invention are: by using an improved octagonal calibration flight scheme, combined with a Transformer-based semi-supervised learning residual error compensation method, the accuracy and efficiency of soft compensation are improved. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart of a soft compensation method for UAV magnetic gradient data based on semi-supervised learning.

[0027] Figure 2 This is a schematic diagram of a soft compensation system for UAV magnetic gradient data based on semi-supervised learning.

[0028] Figure 3 For flight plan design drawings.

[0029] Figure 4 To utilize the Transformer network flowchart.

[0030] Figure 5 To process real-world data graphs using Transformer networks. Detailed Implementation

[0031] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0032] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0033] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0034] Reference Figures 1-5 This is one embodiment of the present invention, which provides a soft compensation method for UAV magnetic gradient data based on semi-supervised learning, including the following steps:

[0035] S1. Design an octagonal calibration flight scheme in which the UAV performs a single maneuver on each flight path to obtain data on the response of the interfering magnetic field to changes in flight attitude.

[0036] S1.1. Centered on a point with a small change in the geomagnetic gradient, select a stable area as the measurement site, set up flight paths with different azimuth angles, form an octagonal calibration flight plan, and perform pitch and roll maneuvers twice to obtain response data of the interference magnetic field to changes in flight attitude.

[0037] Specifically, a relatively stable area was selected as the measurement site. A ground magnetometer was used to measure the magnetic field to determine the area with small gradient changes. Using the center point of the area with small gradient changes as the center, eight flight paths were set with different azimuth angles (such as 270°, 135°, 0°, 45°, 90°, 180°, 225° and 315°, etc.). A total of 16 flight paths were executed. Pitch and roll maneuvers were performed twice, while yaw maneuvers were avoided, so as to comprehensively obtain the response data of the interfering magnetic field to changes in flight attitude.

[0038] Furthermore, the pitch angle varies between 0° and 10°, the roll angle varies between -10° and 10°, and the flight speed is taken as an example value of 10 m / s (the specific value can be modified according to the performance parameters of the aircraft used). It is assumed that the FTMG value of the geomagnetic field on the learning flight trajectory is calculated based on the IGRF model to reduce the difficulty of flight operation, avoid the magnetometer from entering the signal dead zone, improve the consistency of actions and the stability of the compensation coefficient, and adapt to the environment of high wind speed and complex weather.

[0039] S2. Based on the high-altitude compensated flight data and low-altitude survey line flight data in the response data of the interference magnetic field to the flight attitude change, a simulation dynamic measurement noise dataset is established through forward modeling formula and divided into training set, validation set and test set.

[0040] S2.1. The high-altitude compensated flight data in the response data of the interference magnetic field to the change of flight attitude is used as noise-labeled data, and the low-altitude survey line flight data is used as unlabeled data. The mutual conversion between the geographic coordinate system, the geocentric coordinate system and the navigation coordinate system is handled by the coordinate transformation theory formula.

[0041] Specifically, the measured data includes high-altitude compensated flight data and low-altitude survey line flight data. Dynamic noise data (reflecting changes in dynamic noise interference) is obtained from high-altitude four-directional (north, east, south, and west) flight measurements above the study area. Multi-source error data obtained from conventional low-altitude flight measurements is used as unlabeled data. The conversion between geographic coordinate system, geocentric coordinate system, and navigation coordinate system is handled using coordinate transformation theory formulas.

[0042] It should be noted that during actual measurement operations, the coordinates of the measurement area points and the aircraft attitude angle information of the rotary-wing UAV aeromagnetic system are provided by the integrated airborne inertial navigation system (INS). The point coordinates output by the INS are geographic coordinates. The relative deflection of the aircraft attitude angle in the aircraft coordinate system with the magnetic probe as the origin and in the NE-GD coordinate system with the aircraft as the origin is defined. When constructing the error compensation model, it is necessary to design a learning flight and solve for the corresponding compensation coefficients of the UAV. When calculating the normal Earth tensor data based on the IGRF model, it is necessary to differentiate the magnetic potential in the geocentric coordinate system and convert the result to the NE-GD coordinate system. It is necessary to use the geocentric rectangular coordinate system as a bridge to realize the mutual conversion between the geographic coordinate system and the NE-GD coordinate system, and represent the compensation result in the NE-GD coordinate system. Then, the data can be directly used to detect underground targets.

[0043] It should be noted that the expression for the transition from the geographic coordinate system to the geocentric rectangular coordinate system is as follows:

[0044] (1);

[0045] (2);

[0046] (3);

[0047] in:

[0048] (4);

[0049] (5);

[0050] Geographical longitude, Geographical latitude, For the earth to be high, For the semi-major axis of the ellipsoid, For the minor semi-axis of the ellipsoid, Let be the radius of curvature of the circle. The first eccentricity of the Earth's reference ellipsoid.

[0051] From the geocentric rectangular coordinate system to the northeast navigation coordinate system (origin at (…) , , ), corresponding latitude ,longitude )

[0052] (6);

[0053] Where the rotation matrix :

[0054] (7);

[0055] S2.2. Simulation data is generated in batches by exhaustively enumerating the combination of attitude changes and compensation coefficients through forward modeling formulas. The high-altitude compensated flight data and low-altitude survey line flight data in the response data of the interference magnetic field to the flight attitude change are merged with the simulation data to construct a simulation dynamic measurement noise dataset. The simulation dynamic measurement noise dataset is divided into training set, validation set and test set.

[0056] Specifically, based on measured data and forward modeling formulas, a systematic enumeration of various attitude changes and compensation coefficients is conducted to increase the amount of simulation data and enhance the diversity and richness of the simulation dataset, thereby constructing a comprehensive and corresponding simulation training dataset; and the dataset is divided into training set, validation set and test set according to the example values ​​in a ratio of 6:2:2.

[0057] It should be noted that the expression for the forward formula is:

[0058] (8);

[0059] (9);

[0060] (10);

[0061] (11);

[0062] (12);

[0063] (13);

[0064] (14);

[0065] (15);

[0066] in, Yaw angle The pitch angle, R is the roll angle; R is the coordinate transformation coefficient matrix obtained based on the three types of attitude change angles, and formula (9) is the exhaustive full tensor magnetic gradient data obtained from the attitude angles, k ij , l ij (i=x,y,z; j=x,y,z) are compensation coefficients, H 0ij (i=x,y,z; j=x,y,z) is the total tensor magnetic gradient under normal geomagnetic field. (i=x,y,z; j=x,y,z) is the full tensor magnetic gradient in the UAV coordinate system. (i=x,y,z; j=x,y,z) is the full tensor magnetic gradient of the disturbance magnetic field in the UAV coordinate system.

[0067] S3. Construct a semi-supervised learning network model based on Transformer to compensate for the interfering magnetic field in the full tensor magnetic gradient data, and calculate the theoretical full tensor magnetic gradient data based on the international geomagnetic reference field model.

[0068] The S3.1 semi-supervised learning network model based on Transformer includes an input module, a full tensor magnetic gradient encoder, an error-compensated decoder, and an output module.

[0069] Specifically, the full tensor magnetic gradient encoder based on the Transformer semi-supervised learning network model consists of multiple stacked encoding modules, each including a multi-head self-attention layer and a feedforward neural network layer; the error-compensated decoder guides the decoding process through a semi-supervised loss function, establishing a quantitative relationship between the feature vector and the full tensor magnetic gradient noise output sequence.

[0070] S3.2 uses an attention mechanism to extract features from multi-error-source data, compensates for interfering magnetic fields in the full tensor magnetic gradient data, and uses the full tensor magnetic gradient data calculated by the International Geomagnetic Reference Field Model to participate in feature extraction.

[0071] Specifically, based on flight measurement data from multiple error sources using a domestically produced full-tensor magnetic gradient instrument, including high-altitude compensated flight data (used as noise labels for supervised learning) and low-altitude survey line flight data (used as unlabeled data with dimension N), the full-tensor magnetic gradient encoder extracts features from the multi-error-source data of the input module. It consists of multiple stacked encoding modules with identical structures. Each encoding module contains a multi-head self-attention layer and a feedforward neural network layer. It inputs N sets of position-encoded error-source data, obtains Q, K, and V matrices through linear transformation, calculates the correlation between the input data, and extracts latent features. The error-compensated decoder is used for dynamic measurement of gravity gradient noise estimation. The decoding process is guided by a semi-supervised learning loss function to establish a quantitative relationship between the feature vector extracted by the encoder and the desired noise output sequence. The output module outputs dynamic measurement noise, which compensates for the interfering magnetic field in the full-tensor magnetic gradient data. At the same time, it calculates the theoretical full-tensor magnetic gradient data based on the international geomagnetic reference field model. The full-tensor magnetic gradient data calculated based on the international geomagnetic reference field model reflects the normal magnetic field value of the Earth and obtains the magnetic anomaly field caused by underground targets. The normal Earth magnetic field is eliminated simultaneously while compensating for errors.

[0072] S5. Design a multi-constraint loss function.

[0073] S5.1. The multi-constraint loss function is composed of a weighted sum of the label data constraint loss function term, the multi-gradient component transformation relationship constraint loss function term, and the overfitting suppression loss function term.

[0074] Specifically, the expression for the multi-constraint loss function is as follows:

[0075] (16);

[0076] The expression for the label data constraint loss function term is:

[0077] (17);

[0078] in, , This represents the label data from actual high-altitude flight measurements.

[0079] The expression for the loss function term constrained by the multi-gradient component transformation relationship is:

[0080] (18);

[0081] in, , This represents unlabeled data from low-altitude survey lines.

[0082] The expression for the overfit suppression loss function term is:

[0083] (19);

[0084] in, This represents all trainable parameters of the Transformer network.

[0085] The six level components can be converted into each other based on integral relationships as follows:

[0086] (20);

[0087] Formula (20) indicates that the six level components can be converted to each other based on the integral relationship, with superscripts and subscripts indicating the relationship. , , "These correspond to the gravity gradient component data predicted by the Transformer-based semi-supervised learning network model, and the data for high-altitude and low-altitude flight, respectively." , This is a weighting factor used to balance the weights of different loss functions.

[0088] S5.2. The label data constraint loss function is defined based on the difference between the predicted value and the high-altitude compensated flight noise label data. The multi-gradient component transformation relationship constraint loss function is defined based on the quantitative relationship of the integral transformation between the horizontal magnetic gradient components. The overfitting suppression loss function is obtained by balancing the weights of different loss functions in the training of the Transformer semi-supervised learning network model.

[0089] Specifically, The loss function term constrained by the label data is defined by formula (17) and is used to measure the difference between the predicted value and the label data of high-altitude compensated flight noise. The loss function term constrained by the multi-gradient component transformation relationship is defined by formula (18), utilizing the horizontal magnetic gradient component. and The quantitative relationship of integral transformation between them enables the supervision and constraint of labelless low-altitude survey line data; The overfitting suppression loss function term is defined by formula (19) and is used to regularize network parameters to prevent overfitting.

[0090] S6. Using the training and validation sets, optimize the parameters of the Transformer-based semi-supervised learning network model using a stepwise training strategy, and verify the effect based on the test set.

[0091] S6.1. Train the Transformer-based semi-supervised learning network model batch by batch using the training set, and continuously adjust the parameters to minimize the multi-constraint loss function.

[0092] Specifically, a semi-supervised learning network model based on Transformer is trained batch by batch using a simulation training set. The parameters are continuously adjusted to minimize the multi-constraint loss function, including numerical error terms (such as mean squared error, which measures the difference between predicted and true values) and spatial consistency evaluation terms (such as structural similarity index, which measures the spatial smoothness of the prediction results). The Adam optimizer is used, with an initial learning rate of 0.001, a batch size of 16, and a learning rate decay of 50% over 20 training epochs.

[0093] S6.2. Use the validation set to evaluate the performance of the Transformer-based semi-supervised learning network model, calculate the multi-constraint loss value on the validation set, optimize the parameters of the Transformer-based semi-supervised learning network model through cross-validation, and save the optimal parameters of the Transformer-based semi-supervised learning network model with the lowest validation set loss.

[0094] Specifically, during training, the model performance is evaluated using a validation set. The multi-constraint loss value and other metrics (such as accuracy, precision, recall, and F1 score) on the validation set are calculated. The parameters of the Transformer-based semi-supervised learning network model are optimized through cross-validation. If the validation set loss value no longer decreases or starts to increase for five consecutive epochs, training is stopped early, and the optimal parameters of the Transformer-based semi-supervised learning network model with the lowest validation set loss are saved.

[0095] S6.3. Calculate the multi-constraint loss value and performance index on the test set to verify the effectiveness of the semi-supervised learning aeromagnetic interference calibration method.

[0096] Specifically, after the Transformer-based semi-supervised learning network model is trained and validated, the final performance is evaluated using a test set. The multi-constraint loss value and performance index on the test set are calculated, and the effectiveness of the semi-supervised learning aeromagnetic interference calibration method is verified based on the prediction effect on the test set.

[0097] This embodiment also provides a semi-supervised learning-based UAV magnetic gradient data soft compensation system, including: a flight scheme design module, a data acquisition module, a network model construction module, a loss function design module, and a training and verification module; the flight scheme design module is used to design an octagonal calibration flight scheme, in which the UAV performs a single maneuver on each flight path to acquire response data of the interfering magnetic field to changes in flight attitude; the data acquisition module is used to establish a simulated dynamic measurement noise dataset based on high-altitude compensation flight data and low-altitude survey line flight data from the response data of the interfering magnetic field to changes in flight attitude, using forward modeling formulas, and dividing it into training set, verification set, and test set; the network model construction module is used to construct a Transformer-based semi-supervised learning network model to compensate for the interfering magnetic field in the full tensor magnetic gradient data, and calculate the theoretical full tensor magnetic gradient data based on the International Geomagnetic Reference Field Model; the loss function design module is used to design a multi-constraint loss function; the training and verification module is used to optimize the parameters of the Transformer-based semi-supervised learning network model using a stepwise training strategy with the training set and verification set, and verify the effect based on the test set.

[0098] In summary, this invention improves the accuracy and efficiency of soft compensation by providing an improved octagonal calibration flight scheme and combining it with a Transformer-based semi-supervised learning residual error compensation method.

[0099] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for soft compensation of magnetic gradient data of an unmanned aerial vehicle (UAV) based on semi-supervised learning, characterized in that: include, An octagonal calibration flight scheme is designed, in which the UAV performs a single maneuver on each flight path to acquire data on the response of the interfering magnetic field to changes in flight attitude. Based on the response data of the interference magnetic field to the flight attitude change, including high-altitude compensated flight data and low-altitude survey line flight data, a simulation dynamic measurement noise dataset is established using forward modeling formulas and divided into training set, validation set and test set. A semi-supervised learning network model based on Transformer is constructed to compensate for the interfering magnetic field in the full tensor magnetic gradient data, and the theoretical full tensor magnetic gradient data is calculated based on the international geomagnetic reference field model. Design a multi-constraint loss function; Using the training and validation sets, a stepwise training strategy is employed to optimize the parameters of the Transformer-based semi-supervised learning network model, and the performance is validated using the test set. The proposed octagonal calibration flight scheme involves the UAV performing a single maneuver along each flight path. The specific steps are as follows: Centered on points with small changes in geomagnetic gradient, stable areas were selected as measurement sites. Flight routes with different azimuth angles were set up to form an octagonal calibration flight plan. Pitch and roll maneuvers were performed twice to obtain data on the response of the interfering magnetic field to changes in flight attitude. The specific steps for the high-altitude compensated flight data and low-altitude survey line flight data in the response data based on the interference magnetic field to changes in flight attitude are as follows. The high-altitude compensated flight data in the response data of the interference magnetic field to changes in flight attitude are used as noise-labeled data, and the low-altitude survey line flight data are used as unlabeled data. The mutual conversion between the geographic coordinate system, the geocentric coordinate system and the navigation coordinate system is handled by the coordinate transformation theory formula.

2. The method of claim 1, wherein the method is based on semi-supervised learning. The process of establishing a simulated dynamic measurement noise dataset using forward modeling and dividing it into training, validation, and test sets follows these steps. Simulation data is generated in batches by exhaustively combining attitude changes and compensation coefficients using forward modeling formulas. The high-altitude compensated flight data and low-altitude survey line flight data from the response data of the interference magnetic field to flight attitude changes are merged with the simulation data to construct a simulation dynamic measurement noise dataset. The simulation dynamic measurement noise dataset is divided into training set, validation set and test set. 3.The method of claim 1, wherein: The Transformer-based semi-supervised learning network model includes an input module, a full tensor magnetic gradient encoder, an error-compensated decoder, and an output module. By extracting features from multiple error sources through an attention mechanism and compensating for interfering magnetic fields in the full tensor magnetic gradient data, theoretical full tensor magnetic gradient data are calculated based on the international geomagnetic reference field model. 4.The method of claim 1, wherein: The multi-constraint loss function is composed of a weighted sum of the label data constraint loss function term, the multi-gradient component transformation relationship constraint loss function term, and the overfitting suppression loss function term.

5. The method of claim 4, wherein the method is based on semi-supervised learning. The label data constraint loss function is defined based on the difference between the predicted value and the high-altitude compensated flight noise label data. The multi-gradient component transformation relationship constraint loss function is defined based on the quantitative relationship of the integral transformation between the horizontal magnetic gradient components. The overfitting suppression loss function is obtained by balancing the weights of different loss functions in the training of the Transformer semi-supervised learning network model.

6. The method of claim 1, wherein: The training set and the validation set are used to optimize the parameters of the semi-supervised learning network model based on the Transformer by using the step-by-step training strategy, and the effect is verified according to the test set, and the specific steps are as follows, The semi-supervised learning network model based on the Transformer is trained batch by batch using the training set, and the parameters are constantly adjusted to minimize the multi-constraint loss function; The performance of the semi-supervised learning network model based on the Transformer is evaluated using the validation set, the multi-constraint loss value on the validation set is calculated, the parameters of the semi-supervised learning network model based on the Transformer are optimized through cross-validation, and the best semi-supervised learning network model based on the Transformer is saved.

7. The method for soft compensation of UAV magnetic gradient data based on semi-supervised learning as described in claim 1, characterized in that: According to the test set, the effect is verified, and the specific steps are as follows, The multi-constraint loss value and performance indicators on the test set are calculated through the test set to verify the effectiveness of the semi-supervised learning method for aeromagnetic interference calibration.

8. A UAV magnetic gradient data soft compensation system based on semi-supervised learning, based on the UAV magnetic gradient data soft compensation method based on semi-supervised learning in any of claims 1-7, characterized in that: It includes flight plan design module, data acquisition module, network model construction module, loss function design module and training verification module; The flight plan design module is used to design an octagonal calibration flight plan, and the unmanned aerial vehicle performs a single maneuvering action on each flight line to obtain response data of the interference magnetic field to the change of flight attitude. The data acquisition module is used to establish a simulation dynamic measurement noise data set based on the high-altitude compensation flight data and low-altitude flight line flight data in the response data of the interference magnetic field to the change of flight attitude, and divide it into a training set, a validation set and a test set; The network model construction module is used to construct a semi-supervised learning network model based on the Transformer to compensate for the interference magnetic field in the full tensor magnetic gradient data, and calculate the theoretical full tensor magnetic gradient data based on the international geomagnetic reference field model; The loss function design module is used to design a multi-constraint loss function; The training verification module is used to optimize the parameters of the semi-supervised learning network model based on the Transformer by using the training set and the validation set, and the effect is verified according to the test set.

Images