An adaptive weighted geomagnetic navigation deception defense method

Abstract

The present application relates to the technical field of aviation navigation and information security, and particularly relates to a self-adaptive weighted geomagnetic navigation fraud defense method.The present application comprises: acquiring multi-source sensor data of an aircraft, and constructing a theoretical reference geomagnetic field according to a calculated position; extracting first and second perspective features, and performing weighted processing and fusion on the first and second perspective features; obtaining a nonlinear magnetic interference estimation value based on the preliminary weighted fusion features, and correcting geomagnetic measurement data using the nonlinear magnetic interference estimation value; obtaining a geomagnetic navigation positioning result based on the calibrated scalar magnetic field value and the theoretical reference geomagnetic field; comparing a consistency difference value with an adaptive defense threshold value to determine whether satellite navigation signals are subjected to fraud attacks, and switching navigation modes accordingly.The present application improves matching reliability in a region with poor magnetic field characteristics, and realizes accurate suppression and compensation of nonlinear current interference.

Description

Technical Field

[0001] This invention relates to the field of aviation navigation and information security technology, specifically to an adaptive weighted geomagnetic navigation fraud defense method. Background Technology

[0002] Currently, defense against spoofed signals against Global Navigation Satellite Systems (GNSS) primarily relies on signal-level detection technologies. These technologies identify and eliminate anomalous spoofed signals by monitoring changes in the power of the received signal, angle of arrival (AoA) characteristics, or the symmetry of signal correlation peaks. In addition, some high-security systems employ civilian navigation signal protocols based on encryption authentication, aiming to ensure the authenticity of the signal source through identity verification mechanisms.

[0003] A representative advanced technical solution in this field is an end-to-end model based on single-path convolution for feature extraction (see the paper "Gnadt, AR, et al. Signal enhancement for magnetic navigation challenge problem, 2020"). Its benchmark model, Magnav2C, represents the current mainstream level of end-to-end geomagnetic positioning.

[0004] The technical implementation details of this solution are as follows:

[0005] Input feature construction: The system uses the compensated multidimensional geomagnetic components as a one-dimensional time series input. At the input layer, the data is typically standardized, but the significant differences in the characteristics of each component under different geographical environments are not distinguished.

[0006] Single-path convolutional architecture: This model employs a classic deep one-dimensional convolutional neural network. The geomagnetic sequence passes through multiple convolutional layers, pooling layers, and fully connected layers sequentially. The feature extraction process is completed along a single computational path, utilizing the same set of convolutional kernels to perform multi-level abstraction of the geomagnetic signal and extract the statistical fluctuation characteristics of the geomagnetic sequence.

[0007] Static feature fusion: In the fully connected layer stage of the model, the extracted spatial / temporal features are flattened and directly concatenated. The system assumes that each dimension of features contributes equally to the localization result. The network parameters are determined during the training phase through the backpropagation algorithm, but these parameters remain fixed during the online inference phase.

[0008] End-to-end location calculation: The system establishes a direct mapping function from geomagnetic observations to geographic coordinate offsets. The training objective is usually to minimize the root mean square error between predicted and true coordinates. The network is guided to learn through a single loss function and lacks the ability to evaluate matching quality and adjust weights in real time.

[0009] General geomagnetic navigation technology typically includes three components: airborne magnetic compensation, geomagnetic matching, and state updates.

[0010] Airborne magnetic compensation: Classic mathematical models such as Tolles-Lawson (refer to the paper "WE Tolles, JD Lawson, Magnetic compensation of mad equipped aircraft, Airborne Instruments Lab.Inc. 74 (1950) 201–1.") or their enhanced versions are used to model and deduct the aircraft's own maneuvering magnetic interference in real time.

[0011] Matching algorithm: Using geomagnetic feature information, the compensated measurement values ​​are searched for similarity with geomagnetic reference maps stored in the airborne database.

[0012] State estimation: The results of geomagnetic matching are fused with data from the inertial navigation system or other navigation sources using a Kalman filter or an unscented Kalman filter to maintain long-term autonomous positioning capability.

[0013] In the context of modern military confrontation and high-security aviation applications, global navigation satellite systems (GNSS) face increasingly severe threats from electromagnetic interference and deception. Especially in complex battlefield environments or under conditions of strong electromagnetic countermeasures, weak satellite signals are highly susceptible to suppressive interference or deceptive manipulation, leading to navigation failures or even enemy takeover. Therefore, developing geomagnetic navigation technology with passive sensing, no radiation, all-weather operation, and resistance to external tampering has become a key means of building highly survivable autonomous navigation systems in GNSS-denied environments, and also an important benchmark for achieving satellite navigation fraud detection and defense.

[0014] However, despite the progress made in theoretical research of existing end-to-end deep learning geomagnetic navigation technology, its technical system still has the following major shortcomings when facing practical aviation applications, especially in scenarios requiring extremely high accuracy and reliability for satellite navigation fraud prevention:

[0015] 1. Limited Feature Extraction Dimensions Limit Positioning Accuracy in Flat Magnetic Field Zones: Existing convolutional neural network models often employ single-path structures for feature extraction. This approach typically only captures the macroscopic trend features of geomagnetic sequences, easily overlooking minute spatial variations and microscopic fingerprint features of the geomagnetic field within specific geographical areas. In flat areas with small or indistinct magnetic field gradients (i.e., flat magnetic field zones), the feature description capability of a single view is severely insufficient, leading to easily blurred, distorted, or significantly drifting matching results, failing to meet the baseline requirements for high-precision navigation.

[0016] 2. Feature fusion lacks adaptability and has poor environmental robustness. Existing technologies typically employ fixed weighting mechanisms when processing geomagnetic features of different dimensions. However, during actual flight, the signal-to-noise ratio and reliability of different feature components fluctuate in real time due to dynamic changes in flight altitude, intermittent interference from high-power airborne electronic equipment, and random noise from sensors. This fixed weighting mechanism cannot dynamically adjust the contribution of each feature according to the current electromagnetic environment, leading to inaccurate feature extraction under complex electromagnetic interference and poor stability of the positioning results.

[0017] 3. Lack of real-time loss feedback mechanism, hindering self-correction. Most current deep learning navigation models employ open-loop processing, where signals are input and the network directly calculates and outputs coordinates. The model lacks a real-time evaluation and loss feedback mechanism for matching quality. When significant deviations occur in the model's output during calculation, the system cannot perform secondary optimization and adjustment of feature weights through a closed-loop feedback loop. This lack of a mechanism leads to the accumulation of positioning errors over time, making robust operation in long-endurance flight missions difficult.

[0018] 4. Accuracy Bottleneck Prevents Effective Satellite Navigation Fraud Defense: Due to the aforementioned limitations in feature extraction dimensions and dynamic weight adjustment capabilities, the average positioning accuracy of existing geomagnetic navigation technology remains relatively high. However, in satellite navigation fraud defense scenarios, attackers often employ slow, covert, and sophisticated inducement methods. Since the background noise and error range of existing geomagnetic positioning even exceed the early inducement deviation range of fraudulent signals, the system cannot effectively distinguish normal geomagnetic matching fluctuations from genuine satellite navigation fraud, thus losing its value as an independent security benchmark for cross-validation and defense. Summary of the Invention

[0019] The purpose of this invention is to address the problems of existing end-to-end geomagnetic navigation technology, such as weak feature extraction capability in areas with gentle magnetic fields, lack of adaptive weight allocation when facing dynamic current interference from airborne equipment, and insufficient navigation and positioning accuracy as an independent benchmark for effectively detecting precision fraud in satellite navigation. The invention proposes an adaptive weighted geomagnetic navigation fraud defense method.

[0020] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0021] An adaptive weighted geomagnetic navigation fraud defense method includes:

[0022] The system acquires geomagnetic measurement data of the aircraft, current data of airborne equipment, and estimated position at each moment. Based on the estimated position at each moment, a geomagnetic reference map is constructed, and the corresponding theoretical reference geomagnetic field is calculated.

[0023] First-view features are extracted from geomagnetic measurement data and airborne equipment current data, and airborne equipment current data is used as second-view features. The first-view features and second-view features are weighted and fused to obtain preliminary weighted fused features.

[0024] The preliminary weighted fusion features are input into a pre-constructed deep neural network, which outputs a nonlinear magnetic disturbance estimate. The nonlinear magnetic disturbance estimate is then used to correct the geomagnetic measurement data to obtain the calibrated scalar magnetic field value.

[0025] Based on the calibrated scalar magnetic field value and the theoretical reference geomagnetic field, the geomagnetic navigation and positioning results are calculated.

[0026] The consistency difference between geomagnetic navigation positioning results and external satellite navigation positioning results is calculated, and an adaptive defense threshold is set based on the geomagnetic feature gradient of the current flight area obtained from the theoretical reference geomagnetic field.

[0027] Based on the comparison between the consistency difference and the adaptive defense threshold, it is determined whether the satellite navigation signal has been subjected to a fraud attack, and the navigation mode is switched accordingly.

[0028] Furthermore, the geomagnetic reference map is constructed based on the estimated location at each moment, and the corresponding theoretical reference geomagnetic field is calculated, expressed by the formula:

[0029]

[0030] in, Indicates at time The calculated position is The theoretical reference geomagnetic field at that location Longitude Latitude For height; This indicates the coordinate position calculated based on the International Geomagnetic Reference Field Model. And the corresponding time is The Earth's main magnetic field component; This indicates that the coordinates obtained from a pre-stored high-resolution geomagnetic anomaly grid database via bilinear interpolation are: Geomagnetic anomaly at the location; This indicates the time calculated based on observation data from ground-based reference stations. The daily change correction value.

[0031] Furthermore, the extraction of first-view features based on geomagnetic measurement data and airborne equipment current data includes:

[0032] Timestamps are aligned between geomagnetic measurement data and airborne equipment current data, and data within the same time window is extracted.

[0033] High-frequency noise is removed by using a bandpass filter, and samples containing missing or outlier values ​​are eliminated to obtain a preprocessed synchronous dataset.

[0034] The preprocessed synchronous dataset is concatenated along the feature channel dimension and then normalized to obtain first-person perspective features.

[0035] Furthermore, the weighted processing and fusion of the first-viewpoint features and the second-viewpoint features respectively yields preliminary weighted fused features, including:

[0036] The first-view weights are calculated based on the first-view features and are expressed by the following formula:

[0037]

[0038] in, Indicates the first First-view weights for each feature dimension Represents the balance factor. Indicates the first Prior weights for each feature dimension, Indicates the first The standard deviation of each feature dimension within the current time window Indicates the first The standard deviation of each feature dimension within the current time window This represents the total number of dimensions of the first-person perspective features.

[0039] The first-view features are multiplied element-wise with the first-view weights to obtain the first-view weighted features;

[0040] The second-view weights are calculated based on the second-view features and are expressed by the following formula:

[0041]

[0042]

[0043] in, Indicates the first The activity energy index of each interference source Indicates the number of times within the current time window The sampling point of the first sampling point The current value of each interference source This represents the total number of sampling points within the time window. Indicates the first Second-view weights of each interference source This represents the Sigmoid function. This represents the feedback adjustment coefficient. This represents the loss function value calculated from the model output in the current training iteration;

[0044] The second-view features are multiplied element-wise with the second-view weights to obtain the second-view weighted features;

[0045] By fusing the first-view weighted features with the second-view weighted features, a preliminary weighted fused feature is obtained.

[0046] Furthermore, the pre-built deep neural network is constructed using a feedforward multilayer perceptron, including an input layer, one or more hidden layers, and an output layer, wherein the hidden layers use the Swish activation function.

[0047] Furthermore, the geomagnetic measurement data is corrected using the nonlinear magnetic interference estimate to obtain the calibrated scalar magnetic field value, which is expressed by the formula:

[0048]

[0049] in, express The scalar magnetic field value after time calibration. express The cabin scalar total field measurement value at time t. This represents the Tolles-Lawson linear model term. This indicates the number of feature function terms in the Tolles-Lawson model. Indicates the first The regression coefficients corresponding to each characteristic function term express Time of the first A set of characteristic function values ​​related to flight attitude. express The nonlinear magnetic disturbance estimate output by the time-deep neural network.

[0050] Furthermore, the geomagnetic navigation and positioning results calculated based on the calibrated scalar magnetic field value and the theoretical reference geomagnetic field include:

[0051] The magnetic field observation residuals are calculated based on the theoretical reference geomagnetic field and the calibrated scalar magnetic field values ​​at the current moment.

[0052] The drift error is used as a state vector, and a discretized state transition equation is obtained based on the state vector.

[0053] Observation equations are constructed based on magnetic field observation residuals and state transition equations;

[0054] The predicted covariance matrix is ​​calculated based on the state transition matrix, and the Kalman gain is calculated based on the predicted covariance matrix and the observation matrix.

[0055] The optimal error state estimate is calculated based on Kalman gain, magnetic field observation residuals, observation matrix, and state transition matrix.

[0056] The calculated state of the inertial navigation system is corrected by using the optimal error estimate to obtain the geomagnetic navigation and positioning result.

[0057] Furthermore, the step of determining whether the satellite navigation signal has been subjected to a fraud attack based on the comparison result of the consistency difference amount and the adaptive defense threshold, and switching the navigation mode accordingly, includes:

[0058] The consistency difference is compared with the adaptive defense threshold in real time. When the consistency difference is less than or equal to the adaptive defense threshold, the satellite navigation signal is deemed valid and the normal working mode is maintained.

[0059] When the consistency difference exceeds the adaptive defense threshold but the duration does not exceed the set window, the satellite navigation signal is determined to be in a suspected abnormal state. The system continues to maintain satellite navigation input and enters continuous monitoring and alarm preparation mode. When the consistency difference exceeds the adaptive defense threshold and the duration exceeds the set window, the satellite navigation signal is determined to be under fraud attack. The satellite navigation data input is immediately cut off, and the system enters independent geomagnetic navigation defense mode.

[0060] Compared with the prior art, the significant advantages of this invention are:

[0061] (1) It solves the problem of not being able to effectively remove complex dynamic magnetic interference and greatly improves navigation accuracy: This invention uses a dual-view feature extraction strategy and combines the current data of airborne equipment (such as radio, strobe lights, etc.) to solve the problem that traditional Tolles-Lawson models and existing single-view deep learning models are difficult to capture and cancel nonlinear, high-frequency dynamic current magnetic interference. Compared with existing advanced methods (such as Magnav2C), this invention reduces the root mean square error (DRMS) of positioning by about 45.5 meters on the public dataset (SGL2020, Sander Geophysics Limited 2020), which significantly improves the positioning accuracy of airborne geomagnetic navigation in actual combat.

[0062] (2) It overcomes the limitation of fixed feature weight allocation in existing models and enhances the environmental adaptability of the system: This invention uses a hybrid multi-criteria adaptive weighting mechanism (including standard deviation, energy summation and loss feedback) to solve the problem that the compensation effect of existing technologies is reduced due to fixed feature weights when facing changes in flight status or switching of interference sources; by dynamically adjusting the weight of current features according to the loss value during training, the model can adaptively focus on the most important interference source at the current moment, ensuring the navigation stability of the aircraft under different operating conditions.

[0063] (3) Effective fusion of multimodal data was achieved, and the signal-to-noise ratio of magnetic signals was improved: This invention solves the problem of local key features being masked by directly splicing multi-source sensor data by constructing a feature extraction network that includes a global fluctuation perspective and a local independent perspective; This method can simultaneously retain the overall trend features of the magnetic field and the independent interference features of the current, thereby more thoroughly restoring the real geomagnetic field data from the noisy signal, and providing a high-quality data foundation for subsequent high-precision geomagnetic matching. Attached Figure Description

[0064] Figure 1 This is a flowchart of an adaptive weighted geomagnetic navigation fraud prevention method according to the present invention;

[0065] Figure 2 A flowchart for generating preliminary weighted fusion features for this invention;

[0066] Figure 3 This is a flowchart illustrating the process of determining whether a satellite navigation signal has been subjected to a fraud attack, as described in this invention. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this invention clearer, the 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 merely illustrative and not intended to limit the invention.

[0068] like Figure 1 As shown, this invention provides an adaptive weighted geomagnetic navigation fraud prevention method. The specific steps are as follows:

[0069] S1. Data Acquisition and Reference Field Construction.

[0070] Acquire multi-source sensor data during the aircraft's flight, including geomagnetic measurement data (vector magnetic field data, scalar magnetic field data), onboard equipment current data, and the estimated position at each moment; simultaneously construct a geomagnetic reference map based on the estimated position, including the following sub-steps:

[0071] S1.1. Airborne sensor data acquisition:

[0072] During the mission, the onboard data acquisition system synchronously collects the following three sets of data at a set sampling frequency:

[0073] Vector magnetic field data: The three-axis magnetic field components in the aircraft's coordinate system are collected using a vector magnetometer mounted on the aircraft. The vector magnetic field measurement value at time t is :

[0074]

[0075] in, They represent The magnetic flux density in the x, y, and z axes at any given time.

[0076] Scalar magnetic field data: Total magnetic field strength observations in the environment are collected using a scalar magnetometer installed near the aircraft cabin or fuselage, and recorded as... The cabin scalar total field measurement value at time t is The cabin scalar total field measurement value is the original scalar magnetic field observation value before the magnetic compensation processing of this invention, which includes the real geomagnetic field, magnetic interference from the aircraft body, and nonlinear magnetic interference caused by the current of the airborne equipment.

[0077] During the model training phase, when the training platform is equipped with a tail strut scalar magnetometer, it also simultaneously collects tail strut scalar total field data and records it. The scalar total field reference value of the tail strut at time [time] is The tail boom scalar magnetometer is installed at the tail of the aircraft, away from the cabin and main airborne electronic equipment. The scalar total field data it collects is less affected by magnetic interference from the aircraft body and magnetic interference from the current of the airborne equipment. After time synchronization, anomaly removal and necessary basic correction, it serves as a relatively realistic geomagnetic field reference value during the training phase.

[0078] Airborne equipment current data: Real-time synchronous monitoring of current changes in major airborne electronic equipment using current sensors. The system includes... One of the main sources of electromagnetic interference, recorded The vector of current data collected at each moment is :

[0079]

[0080] in, Indicates the first One source of interference Current intensity at time , , This indicates the number of interference sources in the airborne equipment. In this embodiment, four main interference sources are specifically included: radio communication equipment current... strobe lamp current System output power supply current and the current of the air conditioner fan .

[0081] S1.2. Construction of the geomagnetic reference field:

[0082] Based on the current estimated position information (including longitude) provided by the inertial navigation system ,latitude and height Construct a geomagnetic reference map, and find or interpolate the theoretical reference geomagnetic field corresponding to the estimated location and current time in the geomagnetic reference map. To eliminate environmental factor errors, the reference field model is composed of three superimposed parts:

[0083]

[0084] in, Indicates at time The calculated position is The theoretical reference geomagnetic field at that location; This indicates the coordinate position calculated based on the International Geomagnetic Reference Field Model. And the corresponding time is The Earth's main magnetic field component; This indicates that the coordinates obtained from a pre-stored high-resolution geomagnetic anomaly grid database via bilinear interpolation are: Geomagnetic anomaly at the location; This indicates the time calculated based on observation data from ground-based reference stations. The daily change correction value.

[0085] S1.3. Data Alignment and Cleaning:

[0086] The collected vector magnetic field data Total scalar field measurement values ​​of the cabin and onboard equipment current data Perform timestamp alignment and extract identical time windows. The data within the dataset is processed, and a bandpass filter is used to remove high-frequency noise and eliminate samples containing missing values ​​or obvious outliers to obtain a synchronous dataset for feature extraction.

[0087] S2. Dual-view feature extraction.

[0088] A dual-view feature extraction model is constructed to extract features from multi-source sensor data from both a global fusion perspective and a local independent perspective, resulting in first-view features and second-view features. The aim is to capture the joint global trend features of magnetic fields and currents, as well as the independent physical features of currents, through two parallel processing paths, including the following sub-steps:

[0089] S2.1. Constructing a First-Person Perspective: Global Fusion Feature Extraction:

[0090] First, construct the joint feature vector. Vector magnetic field Total scalar field measurement values ​​of the cabin and onboard equipment current data Concatenate along the feature channel dimension:

[0091]

[0092] in, express The joint eigenvector at time step 1.

[0093] Subsequently, the joint eigenvectors are normalized using maximum and minimum values ​​to eliminate dimensional differences. For The first in Each feature dimension has a normalized value. The calculation formula is:

[0094]

[0095] in, These are the first eigenvectors in the joint feature vector. The minimum and maximum values ​​of each feature dimension within the current time window. Indicates the first Each feature dimension The joint eigenvector at time step 1. Indicates the first Each feature dimension The first-person perspective features at any given time are the output multidimensional vectors, which are the first-person perspective features. This feature eliminates the influence of dimensions, enabling subsequent steps to fairly assess the importance of each dimension based on the relative fluctuations of the data.

[0096] S2.2. Constructing a Second Perspective: Local Independent Feature Extraction

[0097] Instead of stitching together magnetic field and current data, the onboard equipment current data is preserved. The independence of [the data]. From this perspective, without normalization, the original physical values ​​are directly used as second-perspective features. :

[0098]

[0099] The original values ​​are retained for subsequent calculations of the physical energy of the current, and the absolute contribution of a specific current value to magnetic field disturbance is directly learned through backpropagation of a neural network.

[0100] S3. Hybrid multi-criteria adaptive weighting.

[0101] like Figure 2 As shown, a hybrid multi-criteria adaptive weighting mechanism is adopted. A standard deviation-based criterion is used to weight the first-view features, and a summation and loss feedback-based criterion is used to weight the second-view features. The mechanism includes the following sub-steps:

[0102] S3.1. First-person weighting: Criterion based on standard deviation:

[0103] First-view features output in step S2.1 The volatility of data can be used to assess the amount of information or the strength of interference it contains.

[0104] First, calculate the standard deviation within the sliding window. For the... Each feature dimension in the time window Calculate the standard deviation of the data sequence within it. :

[0105]

[0106] in, Indicates the first The standard deviation of each feature dimension within the current time window The total number of sampling points within the time window. The first in the current time window The first feature dimension First-view features of each sampling point For the first The mean of each feature dimension within the current time window.

[0107] Then, a balance factor is introduced. To adjust the ratio of prior knowledge to real-time statistical characteristics, combined with preset prior weights. Obtain the first-person perspective weights :

[0108]

[0109] in, Indicates the first First-view weights for each feature dimension Indicates the first Prior weights for each feature dimension, Indicates the first The standard deviation of each feature dimension within the current time window This represents the total number of dimensions for the first-person perspective features. This step enables the model to automatically focus on feature dimensions that fluctuate dramatically.

[0110] The first-view features are multiplied element-wise with the first-view weights to obtain the first-view weighted features.

[0111] S3.2. Second-Perspective Weighting: Based on the Criterion of Summation and Feedback

[0112] Independent second-view features output in step S2.2 A dynamic weighting mechanism combining energy summation and training loss feedback is adopted.

[0113] First, perform an energy summation assessment. Calculate the... One interference source within the time window The sum of the absolute values ​​of the currents within the range is defined as the activity energy index. :

[0114]

[0115] in, Indicates the first The activity energy index of each interference source Indicates the number of times within the current time window The sampling point of the first sampling point The current value of each interference source. This indicator directly reflects the total interference energy of the device during the current time period. This represents the total number of sampling points within the time window.

[0116] Then, a loss feedback mechanism is introduced to calculate the final weights. During model training, the loss feedback mechanism dynamically adjusts the weights of each interference source current characteristic based on the current training error, obtaining the feedback adjustment parameters and weight calculation model after training. During model inference, since a supervised target is unavailable, the current loss function value does not participate in real-time updates. At this point, the feedback adjustment coefficients, loss feedback parameters, and weight calculation model determined during training are fixed. The activity energy index is calculated based on the current data of each interference source within the current time window, and this activity energy index is input into the fixed second-view weight calculation model to obtain the second-view weights for the inference stage. In other words, the loss feedback during training is used to learn and solidify the weight adjustment rules, while the inference stage calculates the weights based on the current current activity and the solidified weight adjustment rules, thus achieving online adaptive weighting under unsupervised labeling conditions. The supervised target is the airborne magnetic interference reference value constructed during training. Specifically, it is the cabin scalar total field measurement value. It originates from a scalar magnetometer installed in the aircraft cabin or near the fuselage, which is... The total magnetic field strength observations are collected directly at all times and have not yet undergone the magnetic compensation processing of this invention. These observations include the actual Earth's magnetic field, magnetic interference from the aircraft body, and nonlinear magnetic interference caused by onboard equipment currents. Tail strut scalar total field reference value. The data is derived from a scalar magnetometer installed on the tail boom of the aircraft during the training phase. Its data acquisition location is far from the cabin and major onboard electronic equipment, thus being less affected by magnetic interference from the aircraft body and the current magnetic interference from onboard equipment. After time synchronization, anomaly removal, and necessary basic calibration, it serves as a relatively realistic geomagnetic field reference value for the training phase. Based on this, the monitoring target... Measurements of the total scalar field in the cabin Reference value of total scalar field of tail strut The difference is determined, that is:

[0117]

[0118] The energy index is nonlinearly modulated using this loss function value to calculate the second-view weights. :

[0119]

[0120] in, Indicates the first Second-view weights of each interference source This is the Sigmoid function, used to map energy to a reasonable range; This is the feedback adjustment coefficient. When the model predicts the loss... When the current is large, the weights of current features with higher energy are automatically increased, thereby forcing the neural network built later to focus on learning complex disturbance patterns that are currently difficult to fit.

[0121] The second-view features are multiplied element-wise with the second-view weights to obtain the second-view weighted features.

[0122] S4. Neural network magnetic compensation and geomagnetic matching.

[0123] The weighted features are input into a deep neural network for magnetic interference compensation, and the calibrated geomagnetic field measurements are output. These measurements are then matched with a geomagnetic reference map to calculate the location observations, including the following sub-steps:

[0124] S4.1. Deep Neural Network Modeling and Compensation:

[0125] Specifically, this involves constructing a deep neural network model to fit the nonlinear magnetic interference of airborne equipment, and outputting an estimate of the nonlinear magnetic interference based on the deep neural network model to compensate and correct the geomagnetic measurement data, including the following:

[0126] A deep neural network is constructed using a feedforward multilayer perceptron. The deep neural network includes at least an input layer, one or more hidden layers, and an output layer. The first-view weighted features output in step S3 are then processed. Weighted features with second perspective Feature concatenation and fusion are performed to obtain preliminary weighted fused features, which are then used as the input vector of the deep neural network.

[0127] The hidden layer employs the Swish activation function to enhance the network's ability to fit complex nonlinear magnetic disturbance relationships and mitigate the dead neuron problem that may occur when using the traditional ReLU activation function; the output layer outputs the estimated nonlinear magnetic disturbance value at the current time step. .

[0128] Preferably, each hidden layer in the deep neural network sequentially performs linear transformation, bias superposition, and nonlinear activation processing on the input features. The calculation process can be expressed as follows:

[0129]

[0130] in, Indicates the first The output features of the hidden layer Indicates the first The output or input features of the layer Indicates the first The weight matrix of the layer, Indicates the first Layer bias terms, This represents the Swish activation function.

[0131] In obtaining the estimate of nonlinear magnetic interference Then, the original scalar magnetic field data were jointly compensated by combining the Tolles-Lawson linear model term to obtain the calibrated scalar magnetic field value. The calculation formula is as follows:

[0132]

[0133] in, express The scalar magnetic field value after time calibration. express The cabin scalar total field measurement value at time t. This represents the Tolles-Lawson linear model term. This indicates the number of feature function terms in the Tolles-Lawson model. Indicates the first The regression coefficients corresponding to each characteristic function term express time A set of characteristic function values ​​related to flight attitude. express The nonlinear magnetic disturbance estimate output by the time-deep neural network.

[0134] In one specific embodiment, a Tolles-Lawson linear model containing 18 feature function terms is used. In other embodiments, Depending on the aircraft platform, magnetometer installation method, and selected... The model order was adjusted.

[0135] By jointly modeling the above linear and nonlinear compensation terms, we can achieve synergistic suppression of aircraft maneuvering magnetic interference and nonlinear magnetic interference caused by airborne equipment current, thereby obtaining a calibrated scalar magnetic field value that is closer to the real geomagnetic field, providing high-precision observation input for subsequent geomagnetic matching and navigation status updates.

[0136] S4.2. Construction of geomagnetic matching observations:

[0137] The calculated position based on the output of the inertial navigation system The location and current time are found or interpolated in the geomagnetic reference map constructed in step S1. Corresponding theoretical reference geomagnetic field Calculate the residuals from the magnetic field observations, which will then be used as input measurements for subsequent filters.

[0138]

[0139] in, Indicates the first Magnetic field observation residuals at each time step; Indicates the first Each time step corresponds to a discrete sampling time; Indicates the first The calibrated scalar magnetic field value corresponding to each time step; Indicates the inertial navigation system in the first... The estimated position output at each time step; Indicates the calculated position and the current moment The theoretical reference geomagnetic field is determined by the geomagnetic reference map and diurnal variation correction.

[0140] S5.EKF (Extended Kalman Filter) navigation state update.

[0141] Based on position observations, the extended Kalman filter algorithm is used to update and correct the state of the inertial navigation system, correcting the drift error of the inertial navigation system and outputting the final navigation and positioning result. This includes the following sub-steps:

[0142] S5.1. Establish the error state equation:

[0143] The drift error of the inertial navigation system is selected as the state vector. Define a 15-dimensional error state vector:

[0144]

[0145] in, This refers to the positional error; For speed error; This refers to attitude error; and These represent the zero-bias errors of the accelerometer and gyroscope, respectively. The discretized state transition equation describes the evolution of the error over time:

[0146]

[0147] in, To make the first The state transitions at time step 1 to the 2nd time step. The state transition matrix at each time step. For the first The process noise input matrix at the time step is used to input the system process noise. Mapped to the error state space; For the first The system process noise at each time step follows a zero-mean Gaussian distribution. For the first The process noise covariance matrix of the time step is used to describe the statistical distribution characteristics of the system process noise and the strength of random disturbances to state components such as position error, velocity error, attitude error and sensor bias error. Indicates the first The state vector at each time step Indicates the first The state vector at each time step.

[0148] S5.2. Establish the observation equation:

[0149] Establish magnetic field observation residuals With state vector The linearization relationship between them. Based on the first-order Taylor series expansion and neglecting higher-order terms, the observation equation is expressed as:

[0150]

[0151] in, The first one calculated in step S4.2 The residual of the magnetic field observation at the time step; For the first Observation noise at time steps For the first The observation noise covariance matrix at each time step; For the first The observation matrix at each time step. It mainly consists of the position gradient of the theoretical reference geomagnetic field at the current calculated location, and is used to describe the influence of position error on the magnetic field observation residual. Since the error state vector selected in this embodiment includes position error, velocity error, attitude error, accelerometer bias error, and gyroscope bias error, and the magnetic field observation residual mainly directly reflects the position error, the observation matrix can be expressed as:

[0152]

[0153]

[0154] in, Indicates the first The observation matrix at each time step; This represents the current location variable, specifically including three location components: longitude, latitude, and altitude. These represent longitude, latitude, and altitude, respectively. Indicates the theoretical reference geomagnetic field Relative to the position variable at the current estimated position Spatial gradient; These represent the partial derivatives of the total theoretical reference geomagnetic field strength with longitude, latitude, and altitude, respectively. This represents a zero vector with one row and three columns, used to indicate that the magnetic field observation residual is not directly sensitive to velocity error, attitude error, accelerometer bias error, and gyroscope bias error.

[0155] S5.3. Filter Update and Closed-Loop Correction:

[0156] Perform the extended Kalman filter iterative steps. First, calculate the state one-step prediction covariance:

[0157]

[0158] in, To predict the covariance matrix, the subscript... Indicates the use of the first Information about the time step for the first Predictions made at time steps Indicates the first The error state estimation covariance matrix updated by filtering at each time step is used to characterize the uncertainty and correlation of the estimation of each error state component at the previous time step.

[0159] Next, calculate the Kalman gain. :

[0160]

[0161] in, Indicates the first Kalman gain at time step.

[0162] Calculate the optimal error state estimate :

[0163]

[0164] in, Indicates the use of the first Time step information for the first The error state predicted in one step is obtained by time step calculation; Indicates the first The optimal error state estimate after time step measurement update.

[0165] Finally, the optimal error estimate is used to correct the calculated state of the inertial navigation system, thus obtaining the geomagnetic navigation and positioning result. :

[0166]

[0167] in, This indicates the compensated geomagnetic navigation and positioning results; Indicates the estimated position of the inertial navigation system; Indicates the estimated value from the optimal error state. The position error estimation component extracted from it.

[0168] S6. Satellite navigation fraud detection and prevention.

[0169] like Figure 3 As shown, the geomagnetic navigation positioning result output in step S5 is used as a reliable benchmark to perform integrity monitoring and fraud prevention on the satellite navigation positioning result output by the airborne satellite navigation receiver, including the following sub-steps:

[0170] S6.1. Construct consistency detection statistics:

[0171] Real-time acquisition of position coordinates (satellite navigation positioning results) output by the satellite navigation receiver. Calculate the satellite navigation positioning results and the geomagnetic navigation positioning results output in step S5. The Euclidean distance deviation between them is used to construct a consistency difference quantity. :

[0172]

[0173] S6.2. Set dynamic defense threshold:

[0174] First, obtain the geomagnetic feature gradient of the current flight area. This gradient is obtained by spatially differentiating the theoretical reference geomagnetic field constructed in step S1.2 at the current calculated location (i.e., the position derivative term of the observation matrix in step S5.2). Then, an adaptive defense threshold is set based on the acquired geomagnetic feature gradient. This threshold is related to the norm of the geomagnetic feature gradient. They are negatively correlated, and the specific calculation formula is as follows:

[0175]

[0176] in, As a preset safety factor, To prevent extremely small values ​​with a denominator of zero, the threshold is automatically tightened when the geomagnetic gradient is large (i.e., the region has abundant magnetic field characteristics) to improve the detection sensitivity for minor fraud biases.

[0177] S6.3. Implement a tiered defense strategy:

[0178] Consistency detection statistics With adaptive defense threshold Perform real-time comparison and execute defense logic: when When the satellite navigation signal is deemed valid, the system maintains normal operating mode; when However, if the duration does not exceed the set window, the satellite navigation signal is determined to be in a suspected abnormal state, the system continues to maintain satellite navigation input, and simultaneously enters continuous monitoring and alarm preparation mode; when If the duration exceeds the set window, it is determined that the satellite navigation signal has been subjected to a fraud attack. The system immediately cuts off the satellite navigation data input and enters the independent geomagnetic navigation defense mode, relying entirely on the geomagnetic navigation positioning results output in step S5 for autonomous navigation.

[0179] This completes the entire process of a single navigation calculation. The system uses the inertial navigation system's calculated state, updated and corrected by extended Kalman filtering in step S5, as the initial state for the navigation calculation in the next time step. The corrected inertial navigation system calculated state includes at least corrected position, velocity, and attitude information. In the next time step, the inertial navigation system continues to recursively calculate based on this corrected navigation state, and combines newly acquired geomagnetic measurement data, onboard equipment current data, and satellite navigation positioning results to iteratively execute steps S1 to S6, achieving high-precision, passive autonomous navigation and real-time defense against satellite deception during long-duration flight.

[0180] To further verify the effectiveness and superiority of the above method, a comparative experiment was conducted based on publicly available airborne geomagnetic measurement datasets.

[0181] I. Experimental Environment Setup.

[0182] The experiment used the SGL2020 public dataset, which contains data collected by the Cessna Caravan aircraft under complex maneuvering flight conditions. Thirteen flight paths exhibiting severe maneuvering characteristics were selected as the test set. Data sources included vector magnetic field data collected by four three-axis vector magnetometers, total magnetic field strength data collected by one scalar magnetometer, and data from the onboard current sensor. The root mean square (DRMS) value of the positioning error was used as the evaluation metric. A smaller DRMS ​​value indicates higher positioning accuracy of the navigation system.

[0183] II. Comparison of experimental results.

[0184] To ensure a fair evaluation, the method of this invention was compared with the following existing mainstream methods under the same testing environment:

[0185] DNN: Ordinary deep neural network; KAN: Kolmogorov-Arnold Network; SAWD-AC: Existing adaptive weighted aeromagnetic compensation method; Magnav2A (M2A), Magnav2B (M2B), Magnav2C (M2C) and Magnav3W (M3W) are typical geomagnetic navigation comparison methods in the MagNav project, among which Magnav2C is a more representative end-to-end geomagnetic navigation benchmark model; Transformer: Deep learning model based on self-attention mechanism; TCN: Temporal Convolutional Network.

[0186] Under the same test conditions, the average positioning error of each method on 13 routes was compared and the results showed that the TL model could only handle linear interference and could not utilize the current data information of the airborne equipment, so it could not cope with complex nonlinear electromagnetic interference and had the largest positioning error. Although the Magnav2C method introduced an attention mechanism, it did not distinguish between global and local perspectives for multi-source data, and the weight allocation mechanism was relatively fixed, which made it less adaptable to high-frequency current fluctuations.

[0187] In contrast, the method of this invention, benefiting from the dual-view feature extraction strategy in step S2 and the hybrid multi-criteria adaptive weighting mechanism in step S3, can more accurately identify and remove dynamic current interference. As shown in Table 1, compared with the existing advanced method Magnav2C, the method of this invention reduces the average DRMS ​​by approximately 45.5 meters on the test route. This strongly demonstrates the significant accuracy advantage and robustness of this invention in complex interference environments.

[0188] Table 1 Comparative Experimental Data

[0189]

[0190] Here, DHAGN represents the airborne geomagnetic navigation method proposed in this paper, which is based on dual-view feature extraction and hybrid multi-criteria adaptive weighting.

[0191] In summary, this invention provides an airborne geomagnetic navigation method based on dual-view feature extraction and hybrid multi-criteria adaptive weighting. By constructing a dual-path feature extraction network that includes global trend and local independence, and designing a hybrid adaptive weighting mechanism that combines standard deviation, energy summation, and loss feedback, it successfully solves the problems of low compensation accuracy and poor adaptability of existing technologies when facing complex dynamic current interference from airborne equipment. Combined with neural network nonlinear compensation and extended Kalman filtering, this invention achieves high-precision passive autonomous navigation.

[0192] It should be noted that the above description is merely a preferred embodiment of the present invention and not a limitation thereof. For those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0193] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. An adaptive weighted geomagnetic navigation fraud prevention method, characterized in that, include: The system acquires geomagnetic measurement data of the aircraft, current data of airborne equipment, and estimated position at each moment. Based on the estimated position at each moment, a geomagnetic reference map is constructed, and the corresponding theoretical reference geomagnetic field is calculated. First-view features are extracted based on geomagnetic measurement data and airborne equipment current data, and airborne equipment current data is used as second-view features. The first-view features and the second-view features are weighted and then fused to obtain preliminary weighted fused features. The preliminary weighted fusion features are input into a pre-constructed deep neural network, which outputs a nonlinear magnetic disturbance estimate. The nonlinear magnetic disturbance estimate is then used to correct the geomagnetic measurement data to obtain the calibrated scalar magnetic field value. Based on the calibrated scalar magnetic field value and the theoretical reference geomagnetic field, the geomagnetic navigation and positioning results are calculated. The consistency difference between geomagnetic navigation positioning results and external satellite navigation positioning results is calculated, and an adaptive defense threshold is set based on the geomagnetic feature gradient of the current flight area obtained from the theoretical reference geomagnetic field. Based on the comparison between the consistency difference and the adaptive defense threshold, it is determined whether the satellite navigation signal has been subjected to a fraud attack, and the navigation mode is switched accordingly.

2. The adaptive weighted geomagnetic navigation fraud prevention method according to claim 1, characterized in that, The process involves constructing a geomagnetic reference map based on the estimated location at each moment and calculating the corresponding theoretical reference geomagnetic field, expressed by the formula: in, Indicates at time The calculated position is The theoretical reference geomagnetic field at that location Longitude Latitude For height; This indicates the coordinate position calculated based on the International Geomagnetic Reference Field Model. And the corresponding time is The Earth's main magnetic field component; This indicates that the coordinates obtained from a pre-stored high-resolution geomagnetic anomaly grid database via bilinear interpolation are: Geomagnetic anomaly at the location; This indicates the time calculated based on observation data from ground-based reference stations. The daily change correction value.

3. The adaptive weighted geomagnetic navigation fraud prevention method according to claim 1, characterized in that, The extraction of first-view features based on geomagnetic measurement data and airborne equipment current data includes: Timestamps are aligned between geomagnetic measurement data and airborne equipment current data, and data within the same time window is extracted. High-frequency noise is removed by using a bandpass filter, and samples containing missing or outlier values ​​are eliminated to obtain a preprocessed synchronous dataset. The preprocessed synchronous dataset is concatenated along the feature channel dimension and then normalized to obtain first-person perspective features.

4. The adaptive weighted geomagnetic navigation fraud prevention method according to claim 1, characterized in that, The first-view feature and the second-view feature are weighted and fused respectively to obtain preliminary weighted fused features, including: The first-view weights are calculated based on the first-view features and are expressed by the following formula: in, Indicates the first First-view weights for each feature dimension Represents the balance factor. Indicates the first Prior weights for each feature dimension. Indicates the first The standard deviation of each feature dimension within the current time window Indicates the first The standard deviation of each feature dimension within the current time window This represents the total number of dimensions of the first-person perspective features. The first-view features are multiplied element-wise with the first-view weights to obtain the first-view weighted features; The second-view weights are calculated based on the second-view features and are expressed by the following formula: in, Indicates the first The activity energy index of each interference source Indicates the number of times within the current time window The sampling point of the first sampling point The current value of each interference source This represents the total number of sampling points within the time window. Indicates the first Second-view weights of each interference source This represents the Sigmoid function. This represents the feedback adjustment coefficient. This represents the loss function value calculated from the model output in the current training iteration; The second-view features are multiplied element-wise with the second-view weights to obtain the second-view weighted features; By fusing the first-view weighted features with the second-view weighted features, a preliminary weighted fused feature is obtained.

5. The adaptive weighted geomagnetic navigation fraud prevention method according to claim 1, characterized in that, The pre-built deep neural network is constructed using a feedforward multilayer perceptron, including an input layer, one or more hidden layers, and an output layer, wherein the hidden layers use the Swish activation function.

6. The adaptive weighted geomagnetic navigation fraud prevention method according to claim 1, characterized in that, The geomagnetic measurement data is corrected using the nonlinear magnetic interference estimate to obtain the calibrated scalar magnetic field value, which is expressed by the formula: in, express The scalar magnetic field value after time calibration. express The cabin scalar total field measurement value at time t. This represents the Tolles-Lawson linear model term. This indicates the number of feature function terms in the Tolles-Lawson model. Indicates the first The regression coefficients corresponding to each characteristic function term express Time of the first A set of characteristic function values ​​related to flight attitude. express The nonlinear magnetic disturbance estimate output by the time-deep neural network.

7. The adaptive weighted geomagnetic navigation fraud prevention method according to claim 1, characterized in that, The geomagnetic navigation and positioning results, calculated based on the calibrated scalar magnetic field value and the theoretical reference geomagnetic field, include: The magnetic field observation residuals are calculated based on the theoretical reference geomagnetic field and the calibrated scalar magnetic field values ​​at the current moment. The drift error is used as a state vector, and a discretized state transition equation is obtained based on the state vector. Observation equations are constructed based on magnetic field observation residuals and state transition equations; The predicted covariance matrix is ​​calculated based on the state transition matrix, and the Kalman gain is calculated based on the predicted covariance matrix and the observation matrix. The optimal error state estimate is calculated based on Kalman gain, magnetic field observation residuals, observation matrix, and state transition matrix. The calculated state of the inertial navigation system is corrected by using the optimal error estimate to obtain the geomagnetic navigation and positioning result.

8. The adaptive weighted geomagnetic navigation fraud prevention method according to claim 1, characterized in that, The step of determining whether the satellite navigation signal has been subjected to a fraud attack based on the comparison result of the consistency difference amount and the adaptive defense threshold, and switching the navigation mode accordingly, includes: The consistency difference is compared with the adaptive defense threshold in real time. When the consistency difference is less than or equal to the adaptive defense threshold, the satellite navigation signal is deemed valid and the normal working mode is maintained. When the consistency difference exceeds the adaptive defense threshold but the duration does not exceed the set window, the satellite navigation signal is determined to be in a suspected abnormal state. The system continues to maintain satellite navigation input and enters continuous monitoring and alarm preparation mode. When the consistency difference exceeds the adaptive defense threshold and the duration exceeds the set window, the satellite navigation signal is determined to be under fraud attack. The satellite navigation data input is immediately cut off, and the system enters independent geomagnetic navigation defense mode.

Images