GNSS interference detection system and method based on generative adversarial network

Abstract

The application discloses a GNSS interference detection system and method based on a generative adversarial network, proposes a generative adversarial network (Seg-GAN) based on mutual information regularization and segmented time series modeling, enhances the relevance of latent variables and generated samples through a multi-class interference signal generation and data enhancement module, generates diversified and realistic interference samples, and improves the generalization ability of the model to unknown interference; meanwhile, a segmented iteration and parallel multi-step prediction strategy is introduced, a long sequence is segmented into subsegments for local feature extraction and global time series modeling, the stability of gradient propagation is optimized, the generator generates pseudo samples conforming to the real GNSS signal time series features, and the discriminator accurately captures the local time series feature difference. The application significantly improves the precision and robustness of interference detection, reduces the computational complexity, and is suitable for real-time deployment of lightweight devices such as unmanned aerial vehicles.

Description

Technical Field

[0001] This invention belongs to the field of communication technology and relates to Global Navigation Satellite System (GNSS) interference detection technology, and in particular to a GNSS interference detection system and method based on generative adversarial networks. Background Technology

[0002] Global Navigation Satellite System (GNSS) is the core navigation and positioning method for UAV collaborative sensing systems, and the security and reliability of its signals directly affect mission execution. However, GNSS signals are vulnerable to complex interference attacks such as deception, jamming, and suppression. Among these, deception interference, due to its strong concealment and high similarity to real signals, has become a major factor threatening UAV safety. Various detection schemes for GNSS signal interference have been developed in the industry, but existing schemes have several limitations:

[0003] 1. Traditional interference detection techniques mainly rely on supervised learning methods, using models such as recurrent neural networks (RNN) and long short-term memory networks (LSTM) to extract temporal features. However, supervised learning algorithms are overly dependent on data completeness, and their performance is limited by the consistency of the distribution of training and test data. In unknown interference scenarios not covered by the training data, their generalization ability drops significantly, and their detection performance declines sharply.

[0004] 2. Furthermore, as long-sequence temporal data, GNSS signals present challenges for traditional RNN models, such as gradient explosion and feature capture capability decaying with sequence length. These issues make it difficult for the models to capture the global temporal dependencies of GNSS signals. While Transformer-based models can handle long-range dependencies, their computational complexity is proportional to the square of the sequence length, making them difficult to deploy on resource-constrained UAV embedded platforms.

[0005] 3. While existing Generative Adversarial Networks (GANs) have some applications in signal generation tasks, GNSS interference signals typically exhibit diverse category characteristics (such as frequency-locked interference and carrier-aligned interference) and long-term sequential dependencies. Standard GAN generators often demonstrate highly coupled noise processing methods when handling such tasks, failing to achieve precise control over the semantic features of the generated data by adjusting specific dimensions of the noise input. The generation process lacks category controllability, resulting in insufficient diversity and realism of the generated interference samples. It cannot accurately simulate real interference patterns (such as frequency locking, carrier alignment, and code phase propagation), leading to limited data augmentation effects and difficulty in effectively supporting the robust training of detection models. Furthermore, standard GANs have limited ability to capture long-term temporal dependencies, and the temporal coherence of the generated interference signal samples is insufficient to meet practical requirements, thus limiting the model's generalization performance in complex interference scenarios. Simultaneously, the scarcity and uneven distribution of interference samples in real-world environments further restricts the adaptability of traditional methods.

[0006] 4. Existing methods are not good at capturing local temporal features in dynamic environments and lack robust processing mechanisms for uneven class distribution, resulting in high false positive and false negative rates.

[0007] Therefore, there is an urgent need for an innovative technical solution that can achieve high-precision, low-latency interference detection in complex and unknown environments, while also taking into account the lightweight nature of the model and its real-time performance. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention provides a comprehensive solution that integrates generative adversarial network optimization, lightweight temporal modeling, and dynamic interference determination to improve the anti-interference capability and navigation safety of UAVs in complex electromagnetic environments.

[0009] This invention provides an improved generative adversarial network framework for GNSS interference detection, named Seg-GAN. First, a conditional generation mechanism is introduced into the generator design, decomposing the input into random noise and latent coding to enhance the controllability and diversity of the generated data. Second, mutual information regularization constraints are employed to ensure that the generator can generate corresponding signal samples according to the specified interference category, further improving the accuracy of category control. Finally, considering the long-term temporal characteristics of GNSS signals, a sequence optimization module is designed. This module optimizes the long-sequence modeling capability through piecewise iteration and parallel multi-step prediction strategies, thereby reducing computational complexity while maintaining temporal consistency. By introducing a conditional generation mechanism, mutual information regularization constraints, and piecewise iteration and parallel multi-step prediction techniques, this invention significantly enhances the diversity and category controllability of the generated data, while optimizing the modeling capability of long-sequence signals, thus providing an efficient and robust solution for GNSS interference detection.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] A GNSS interference detection system based on generative adversarial networks (GANs) includes a GAN module; the GAN module includes a generator, a discriminator, and a mutual information regularization module, wherein the generator is used to introduce latent coding... With random noise The composite input is used to generate multi-class interference signal samples that conform to the temporal characteristics of real GNSS signals through a sequence optimization module. The discriminator is used to extract local temporal features through the sequence optimization module and compare them with the real signals, outputting the probability that the signal is a real signal. The mutual information regularization module is used to maximize the latent coding. The mutual information constraint generation process between the generator and the generated signal.

[0012] Furthermore, the generator is used to segment the input noise sequence into fixed-length segments, merge each noise sequence with the category vector to form a composite input subsequence, and send it to the sequence optimization module. The GRU unit performs segmented feature extraction to generate the corresponding hidden state sequence, and converges it into the hidden state of the last segment as the initial input for parallel multi-step prediction. The parallel mechanism is used to generate the output of multiple subsequences simultaneously. After dimensionality transformation and splicing, a pseudo-signal sample sequence with the same size as the real sample is formed.

[0013] Furthermore, the parallel multi-step prediction introduces position coding, which is constructed by concatenating relative position coding and channel position coding.

[0014] Furthermore, the discriminator divides the complete long time-series signal into multiple fixed-length segments and inputs them into the sequence optimization module. The GRU unit extracts the temporal features of each segment, generates the corresponding hidden state sequence, and summarizes them into a feature vector. Based on this feature vector, a classification head performs single-step detection and outputs the probability that the signal is the real signal.

[0015] Furthermore, the hidden state of the previous segment is passed to the next segment as context information.

[0016] Furthermore, the mutual information regularization module includes an auxiliary network. The auxiliary network shares the feature extraction layer with the discriminator and outputs the conditional distribution. During training, the latent encoding is maximized through variational lower bound estimation. Mutual information with the generated samples.

[0017] Furthermore, it also includes an interference determination module, which dynamically adjusts the positioning strategy based on the probability value output by the discriminator.

[0018] Furthermore, it also includes a data preprocessing module for performing in-phase component processing on the input signal. Orthogonal components The extraction and normalization process is calculated using the following formula: , The original signal is then scaled to the [0,1] interval using minimum-maximum normalization.

[0019] This invention also provides a GNSS interference detection method based on generative adversarial networks, comprising the following steps:

[0020] Step 1: Collect training data and preprocess it;

[0021] Step 2: Train the generative adversarial network module in the GNSS interference detection system based on generative adversarial network. The generator generates simulated interference signals based on random noise and latent coding. The preprocessed real samples and simulated interference signals are input together into the discriminator to carry out adversarial training until the training is completed.

[0022] Step 3: During real-time detection, the data is preprocessed and then input into the trained discriminator. The discriminator outputs the interference probability value to the interference determination module, which dynamically adjusts the positioning strategy based on the probability value output by the discriminator.

[0023] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0024] 1. This invention utilizes a multi-class interference signal generation module based on mutual information regularization. By leveraging the strong correlation between latent encoding and generated samples, it generates diverse and realistic interference samples. This overcomes the strong dependence of traditional supervised learning on the identical distribution of training and testing data, maintaining stable detection performance even when faced with unseen complex interference patterns. Through leave-one-out cross-validation, the model maintains a high true positive rate even after removing data from specific scenarios, effectively solving the generalization bottleneck in unknown interference scenarios. It adapts to various complex interference types, significantly improving generalization ability against unknown interference and demonstrating outstanding robustness against interference.

[0025] 2. This invention innovatively employs a segmented iterative and parallel multi-step prediction strategy, dividing long sequences into fixed-length segments for parallel processing. Combined with a lightweight gated recurrent unit (GRU) to capture local temporal dependencies, it significantly reduces the modeling complexity of traditional recurrent networks and effectively alleviates the gradient explosion or vanishing problem. Simultaneously, a bridging mechanism transmits temporal information between segments, maintaining the continuity of global features. The generated pseudo-samples more closely resemble the temporal characteristics of real GNSS signals, allowing the discriminator to accurately capture local temporal differences and improve the feature modeling accuracy of long sequence signals.

[0026] 3. This invention significantly reduces the number of model parameters and computational overhead by sharing a feature extraction layer between the discriminator and the auxiliary network Q. It employs a lightweight sequence optimization module, SegRNN, to avoid redundant computations in highly complex models, improving inference efficiency while maintaining detection performance. This allows for flexible deployment on UAV-borne embedded devices. Gradient penalty and learning rate decay strategies are introduced during training, eliminating the need for high-performance computing equipment and meeting the real-time detection requirements of resource-constrained scenarios.

[0027] 4. This invention employs the Wasserstein distance-optimized adversarial loss function, combined with class-balanced cross-entropy loss to alleviate the problem of uneven data distribution. This enables the model to accurately distinguish between real and interference signals even in complex interference scenarios, effectively reducing false positive and false negative rates. The model can quickly respond to interference signals, providing timely protection for UAV navigation safety. It maintains high detection performance in both static and dynamic interference scenarios, exhibiting high detection accuracy, strong real-time performance, and low false alarm rate.

[0028] 5. This invention simulates complex channel environments by adding Gaussian white noise, and the model maintains stable performance under different interference scenarios. It can achieve efficient detection for various typical interference attack modes, especially in complex scenarios where the interference signal is highly similar to the real signal, and can still maintain reliable detection results. It is fully adaptable to the actual needs of complex electromagnetic interference scenarios such as military reconnaissance and emergency response, and is suitable for a variety of anti-interference scenarios. Attached Figure Description

[0029] Figure 1 This is the architecture diagram of the entire system.

[0030] Figure 2 This is an example diagram of the Seg-GAN model.

[0031] Figure 3 This is a schematic diagram of the sequence optimization module, Seg-RNN.

[0032] Figure 4 This is a schematic diagram of a parallel multi-step prediction module.

[0033] Figure 5 This is a schematic diagram of the interference control software configuration interface.

[0034] Figure 6 This is a schematic diagram of the output from the drone's real-time monitoring terminal. Detailed Implementation

[0035] The technical solutions provided by the present invention will be described in detail below with reference to specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0036] Figure 1 This diagram illustrates the overall system architecture, showing how multiple functional modules work together to achieve precise positioning and stable flight. The core modules include the UAV sensor module, GNSS interference detection module, positioning module, and flight control module. The UAV sensor module is responsible for real-time acquisition of GNSS signals, environmental perception data, and hardware status information. It interacts with external devices via the communication module, and the data processing unit standardizes and integrates the raw data, ultimately forming a unified input source through data aggregation.

[0037] The GNSS interference detection module is based on an improved generative adversarial network (Seg-GAN). It preprocesses the input signal, extracts in-phase (I) and quadrature (Q) components, and calculates key indicators such as carrier-to-noise ratio, amplitude ratio and phase difference through a feature extraction unit, and finally outputs the interference probability value.

[0038] The positioning module dynamically assesses positioning reliability based on interference detection results. In interference-free scenarios, it relies on standard GPS positioning. When the detected interference probability exceeds a threshold, it switches to a multi-source fusion positioning mode (such as visual odometry or inertial navigation) to ensure navigation continuity. Specifically, it includes a GPS positioning module and an interference determination module.

[0039] The flight control module adjusts the UAV's three-dimensional spatial position and flight attitude in real time through pose control, attitude control, and motor control unit. It maintains stability by combining gyroscope and accelerometer data, and precisely adjusts the motor speed according to control commands to achieve dynamic optimization of heading and altitude.

[0040] The entire system achieves functional decoupling through modular design, supports flexible expansion and lightweight deployment, and significantly improves its adaptability to complex interference scenarios such as deception and blocking by relying on the temporal modeling capabilities and data augmentation strategies of Seg-GAN. Furthermore, it ensures the mission continuity and flight safety of UAVs in interference environments through low-latency detection (<50ms) and multi-source positioning switching mechanisms.

[0041] Specifically, the GNSS interference detection system based on generative adversarial networks provided by this invention should at least include the aforementioned GNSS interference detection module, and may further include an interference determination module and other modules in the system. The GNSS interference detection module includes: a data preprocessing module and a generative adversarial network module.

[0042] The data preprocessing module performs in-phase component processing on the input signal. Orthogonal components The extraction and normalization process is calculated using the following formula: , The original signal is then minimized and scaled to the [0,1] interval using a minimum-maximum normalization method; optional Gaussian white noise is added to simulate complex channel environments. The real I / Q signals output by the data preprocessing module are used as real samples and fed into the discriminator for adversarial training. During the real-time online inference phase of the UAV, the GNSS signals collected on-site are processed by the data preprocessing module and then input into the trained discriminator for interference discrimination.

[0043] Generative Adversarial Network (GAN) modules employ methods such as Figure 2 The Seg-GAN framework implementation shown includes a generator, a discriminator, and a mutual information regularization module.

[0044] The interference detection module is used to dynamically adjust the positioning strategy based on the probability value output by the discriminator. Specifically, the interference detection module triggers positioning mode switching according to the output probability threshold, and activates visual inertial odometry-assisted positioning when the interference probability exceeds the set threshold.

[0045] The generator's input consists of two parts: one is random noise. It follows a standard normal distribution. First, it is used to capture the basic statistical characteristics of the generated data; second, it is latent encoding. Latent encoding is composed of several latent variables that follow a specific prior probability distribution (such as a uniform or categorical distribution) and is used to characterize the categorical information of the generated data. The introduction of this feature enables the generator to generate corresponding samples based on the specified type of interference signal (such as frequency-locked interference signal, carrier-aligned interference signal, etc.), thereby achieving category controllability.

[0046] The generator samples random noise from a predefined noise distribution. and with latent encoding Combine them to form a composite input. The composite input is then fed into the generator network. The feature transformation and mapping are performed through a series of neural network layers (including but not limited to fully connected layers, convolutional layers, or recurrent neural units), and the pseudo signal samples are finally output. To ensure that the generated samples are highly consistent with real GNSS interference signals in terms of timing characteristics, this invention integrates a sequence optimization module into the generator to further optimize the generation quality of long sequence signals.

[0047] Compared to traditional GANs, this invention introduces latent encoding. This achieves explicit control over the semantic features of the generated data. However, simply introducing latent encoding is insufficient to guarantee that the generator can effectively utilize this conditional information. During unconstrained training, the generator may degenerate into ignoring... The impact, i.e., the distribution of the generated samples and Equivalence leads to generated samples being irrelevant to the specified category. To address this issue, this invention further introduces mutual information regularization constraints, which optimize the latent encoding. With generated data Mutual information between them enhances the effect of condition generation.

[0048] Mutual information regularization is one of the core innovations of the Seg-GAN framework in this invention. Its purpose is to maximize the relationship between the latent encoding c and the generated data. Mutual information between To ensure that the generator can be based on It generates corresponding interference signal samples according to the specified category, ensuring the controllability of the category of generated samples. In information theory, mutual information... Defined as the reduction in uncertainty about one variable when another variable is known, its mathematical expression is:

[0049]

[0050] in, Representing variables entropy, For a given The conditional entropy after that. In this invention, the optimization objective is set as:

[0051]

[0052] Achieving this optimization objective has significant technical implications. By maximizing... Generate data Able to fully reflect the potential coding The category information it represents. For example, when When specified as a particular type of interference signal, the generator will produce samples highly consistent with that type, rather than random, meaningless signals. Theoretically, when mutual information reaches its maximum value, given generated samples... Then, latent encoding conditional entropy It will tend to the minimum, that is The uncertainty is significantly reduced, thereby enabling the accurate transmission of category information.

[0053] However, directly calculating mutual information It presents significant technical challenges in high-dimensional data scenarios because it involves posterior probabilities. A precise estimate of the posterior probability is difficult to obtain analytically. Therefore, this invention employs a variational information maximization method, introducing an auxiliary distribution... Approximate true posterior distribution The mutual information is then fitted with a lower bound. The specific derivation process is as follows:

[0054]

[0055] Due to direct sampling Since computation is not feasible, this invention utilizes a variational method to introduce an auxiliary distribution. Furthermore, based on the nonnegativity of KL divergence, a lower bound for mutual information is derived:

[0056]

[0057] Based on the above derivation, this invention defines a variational lower bound. for:

[0058]

[0059] in The variable is a constant because the prior distribution of the latent encoding c is fixed during design. Ultimately, the training objective function of Seg-GAN can be expressed as:

[0060]

[0061] in, For the adversarial loss of standard GAN, The hyperparameters for adjusting the strength of mutual information regularization are optimized. When auxiliary distribution Approaching the true posterior At this time, the variational lower bound tends to tighten, thus realizing latent encoding. Maximizing the mutual information between the generated samples and the specified interference type ensures a strong correlation between the generated samples and the specified interference type in terms of carrier phase and power characteristics. This method not only theoretically guarantees the effectiveness of category control but also has strong operability in computational implementation.

[0062] The generative adversarial network module uses Wasserstein distance to optimize the adversarial loss function, and the discriminator loss includes a gradient penalty term.

[0063]

[0064] The generator loss function is:

[0065]

[0066] Combined with mutual information loss Achieve stable training.

[0067] Furthermore, the discriminator integrates a class-balanced cross-entropy loss function:

[0068]

[0069] Class weight It is inversely proportional to the sample frequency and is used to mitigate detection bias caused by uneven data distribution.

[0070] In the specific implementation of the mutual information regularization module, this invention designs an auxiliary network. Used to estimate conditional distribution To improve computational efficiency and enhance feature consistency, Network and Discriminator Some parameters are shared. Specifically, and Shared feature extraction layers (such as sequence optimization modules) are used to reuse low-level signal feature representations. Based on these shared features, Output via an additional fully connected layer ,and Output the true / false probability. This parameter-sharing design has several advantages: First, it reduces the number of model parameters, lowering computational overhead; second, it promotes... and Consistency in feature extraction improves the synergy between category estimation and true / false detection. In GNSS interference detection tasks, the parameter sharing strategy ensures... and We can learn these key patterns together.

[0071] for In modeling, this invention selects a suitable distribution form based on the type of the latent code c. If... If it is a discrete variable (such as the category label of the interference signal), then The class probability is output using the softmax function; if For continuous variables (such as signal strength or offset), the model is a Gaussian distribution, which is derived from... Output the mean and variance parameters. During training, the generator... and auxiliary network Through joint optimization To achieve collaborative work, G maximizes Enhance the correlation between c and the generated data, and Improve the accuracy by minimizing prediction error The estimation accuracy is improved to ensure that the generator can accurately predict the given values. Generate interference signal categories that match expectations.

[0072] This invention significantly improves the controllability and diversity of generated data categories through a conditional generation mechanism and mutual information regularization constraints, providing high-quality pseudo-signal samples for GNSS interference detection tasks. The design and implementation of the sequence optimization module, as well as the specific optimization strategies of the generator and discriminator in long sequence modeling, will be further elaborated below to fully demonstrate the technical advantages of the Seg-GAN framework. This sequence optimization module, using segmented iteration and parallel multi-step prediction as its core strategies, optimizes the feature extraction and generation process of long sequence signals, thereby significantly improving the accuracy and robustness of interference detection. The specific implementation of the sequence optimization module is described in detail below.

[0073] like Figure 3The diagram shows the SegRNN sequence optimization module, which uses segmented iteration and parallel multi-step prediction as its core strategies to optimize the feature extraction and generation process of long sequences. Specifically, this module divides the input GNSS signal sequence into multiple fixed-length segments. GRU units then perform local feature extraction and prediction within each segment, significantly reducing the number of traditional loop calculations and thus improving training stability and computational efficiency. In the segmentation design, the total sequence length is... input sequence Divided into several windows of size The sub-segments, such that the number of sub-segments By employing this segmented processing strategy, the modeling complexity is reduced from that of traditional recurrent networks. Down to This effectively reduces the computational burden. This segmentation strategy, by decomposing long sequences into manageable local segments, avoids the gradient propagation problem on extremely long paths, effectively mitigating gradient explosion or vanishing phenomena. Simultaneously, to ensure sufficient transmission of temporal information between segments, the module designs a bridging mechanism at the end of each segment, passing the hidden state of the previous segment as context information to the next segment. This bridging mechanism maintains the continuity of global temporal features based on local feature extraction. This mechanism not only guarantees the effectiveness of local feature extraction but also maintains the overall consistency of the entire sequence.

[0074] The sequence optimization module is based on a gated recurrent unit (GRU) design, while also supporting alternative implementations of other recurrent neural network units such as Long Short-Term Memory (LSTM) networks. Taking the GRU as an example, it updates the gate... and reset door The specific calculation process for controlling the flow of information and the retention of memory is as follows: First, through... and Calculate the update gate and reset gate; then, based on , Generate candidate hidden states; finally, through Update the hidden state. In practical applications, the input sequence... First, the dimensions are adjusted through a common linear function layer to meet the input requirements of the GRU unit. Then, the hidden state sequences are sequentially input into the GRU unit in a segmented manner to generate the corresponding hidden state sequences. The hidden state sequence contains the temporal features of the input sequence, providing a foundation for subsequent prediction and generation.

[0075] Both the generator and discriminator are implemented using the sequence optimization module, but they differ in functionality and implementation details. The generator's main goal is to generate pseudo-signal samples that are highly consistent with the timing characteristics of real GNSS signals based on the input noise. Its implementation process includes: first, sampling multiple noise sequences from a predefined noise distribution. and divide it into Segment; then, pair each noise sequence with a class vector. Merge to form a composite input subsequence , where the category vector The target signal's specific attributes (such as interference patterns) are encoded to provide constraints for the generation process. These composite subsequences are then fed into a sequence optimization module, where segmented feature extraction is performed using GRU units to generate corresponding hidden state sequences, which are then aggregated into the hidden state of the last segment. Finally, with As the initial input for parallel multi-step prediction, combined with location embedding information, multiple sub-sequences are generated simultaneously using a parallel mechanism. After dimensionality transformation and concatenation, a result identical to the real sample is formed. Pseudo-signal sample sequences of uniform size This process, through the synergy of segmented processing and parallel prediction, not only improves generation efficiency but also ensures the overall consistency of the pseudo-signal over long time.

[0076] The discriminator is responsible for evaluating the authenticity of the input signal; its input is a real sample. Or pseudo-samples generated by the generator The discriminator implementation process includes: first, splitting the complete long time-series signal into multiple fixed-length subsequences; then, inputting these subsequences into the sequence optimization module, extracting the temporal features of each subsequence through the GRU unit, generating the corresponding hidden state sequence, and summarizing them into a feature vector. Finally, based on this feature vector A single-step detection is performed through a classification head (such as a fully connected layer), and the output signal is the probability that it is a true signal. Unlike generators, which focus on global sequence generation, discriminators prioritize feature extraction to capture local temporal characteristics to support subsequent authenticity determination, thereby achieving accurate differentiation between real and fake samples.

[0077] This invention employs a channel independent (CI) strategy when handling multivariate prediction tasks. Specifically, for a multivariate prediction problem involving N variables, this invention transforms it into... An independent univariate prediction problem, namely ,in Compared to traditional methods that use all historical variables for simultaneous prediction, the channel-independent strategy can more efficiently capture the potential relationships between variables, thereby improving prediction performance. This strategy is particularly suitable for GNSS interference detection tasks because it can decompose complex multivariate time-series data into more easily processed single-variable sub-tasks, thus improving the overall detection accuracy of the system.

[0078] like Figure 4 The diagram illustrates a parallel multi-step prediction strategy. Traditional iterative multi-step prediction (RMF) can preserve short-term memory information of the time series, but this leads to error accumulation during the prediction phase, resulting in inaccuracies. This will lead to Furthermore, the accuracy becomes even less accurate, leading to lower accuracy for predictions over longer time spans. Therefore, this invention abandons the traditional iterative multi-step prediction method and adopts a parallel multi-step prediction (PMF) strategy. By using parallel output, the extracted feature information is used simultaneously to predict values ​​at multiple time points or time periods, thereby effectively avoiding the problem of error accumulation.

[0079] Figure 4 It is a parallel output method that predicts a span of... The long sequence prediction task is decomposed into A length of The subsequence prediction problem outputs to multiple output neurons in parallel, simultaneously generating multiple subsequence outputs to predict values ​​at different time points or time periods. This is achieved by introducing [a specific mechanism / mechanism] under the output neurons. It is a positional code used to reflect the sequential order of each output sequence. Encoded by relative position and channel position coding The cascaded construction involves relative position coding reflecting the position of each predicted segment within the complete sequence, while channel position coding identifies the channel position of the current sequence, thus ensuring the temporal accuracy and consistency of the prediction results. In this way, relative position coding and channel position coding are introduced into each subsequence prediction process, avoiding error accumulation in iterative prediction.

[0080] Based on the GNSS interference detection system based on generative adversarial networks, this invention also provides a GNSS interference detection method based on generative adversarial networks, comprising the following steps:

[0081] Step 1: Obtain the TEXBAT dataset as the original training data. The dataset covers the static interference-free scenario DS0, the dynamic interference-free scenario DS1, and the DS2 to DS7 scenarios corresponding to three types of deception interference: code phase propagation, frequency locking, and carrier alignment. Extract the I / Q components from the GNSS radio frequency signals in the dataset, and complete the time series preprocessing by min-max normalization and sliding window stacking to generate standardized time series samples that can be used in the model.

[0082] Step 2: Construct a Seg-GAN model comprising a generator, discriminator, and auxiliary Q-network. The generator and discriminator share a sequence optimization module with a configuration segment bridging mechanism, and are jointly optimized using WGAN-GP adversarial loss combined with mutual information regularization loss. Gaussian distributed random noise vectors and latent codes of interference categories are input into the generator to generate simulated interference signals. Preprocessed real samples and generated simulated samples are input into the discriminator for adversarial training. Leave-one-out cross-validation is used during training. Under the constraint of removing unknown interference scenarios with DS7 carrier alignment and FPR=1%, the model's TPR can reach 90.86%. The discriminator is lightweighted and pruned based on the sequence optimization segmentation structure. After the model is trained offline, only the lightweight discriminator is deployed to an embedded device on a UAV, while the generator and auxiliary Q-network remain on a cloud device.

[0083] Step 3: The UAV's onboard terminal collects GPS and IMU sensor data in real time, reuses the unified preprocessing workflow from Step 1 to obtain temporal features, and directly inputs them into the local lightweight discriminator. The discriminator outputs the probability value of GNSS spoofing interference and transmits it to the interference determination module. The interference determination module judges the interference status according to a preset probability threshold. When there is no interference, GPS positioning is maintained; when spoofing interference is detected, the redundant positioning mode of UWB and vision fusion is switched. The onboard embedded device can periodically upload the preprocessed temporal features to the cloud. The cloud uses a complete Seg-GAN network to complete signal adversarial verification and sends the updated lightweight discrimination parameters to the onboard terminal to realize local model iterative updates.

[0084] like Figure 5 The image shows the interference control software configuration interface. After the communication test passes successfully, start the interference control software and log in to the system. Place the receiving antenna in an open environment to obtain stable GPS ephemeris data. After the ephemeris search is complete, enter the interference source software configuration interface as shown below. Figure 5 As shown, the jammer's transmit power was set, and the interference signal strength was precisely calibrated using a spectrum analyzer. The interference-to-signal ratio was initially set to 20 dB to simulate a moderate-intensity interference scenario; this choice was intended to test the sensitivity of the detection algorithm under typical interference conditions.

[0085] like Figure 6The diagram shows the output of the UAV real-time detection terminal. During the testing phase, the `detect` script was run to achieve real-time detection of GNSS spoofing interference. This script reads sensor data in real time through ROS topics and calculates the signal confidence based on a pre-trained model. Under normal conditions without interference, the model outputs a confidence level close to 1.0. Subsequently, interference was enabled through the interference source software control platform, with the output power set to 20W and the interference mode set to spoofing mode. The programmable attenuation was adjusted using a sliding progress bar to test the algorithm's performance under different interference-to-signal ratios. In the trajectory settings, the trajectory and latitude / longitude coordinates of the target deviation can be selected. After enabling interference, if a spoofing signal is detected, the model output confidence level drops rapidly to close to 0.0, triggering an alarm mechanism. In the real-machine experiment, the interference source transmission power was set to 20W, and two types of interference-to-signal ratio spoofing interference environments of 10dB and 20dB were obtained through programmable attenuation adjustment. The average detection delay in the static trajectory spoofing experiment was 5.50s, and the average detection delay in the dynamic trajectory spoofing experiment was 4.92s. The overall real-machine detection success rate reached 97.22%.

[0086] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A GNSS interference detection system based on generative adversarial networks, characterized in that, It includes a generative adversarial network (GAN) module; the GAN module includes a generator, a discriminator, and a mutual information regularization module, wherein the generator is used to introduce latent coding... With random noise The composite input is used to generate multi-class interference signal samples that conform to the temporal characteristics of real GNSS signals through a sequence optimization module. The discriminator is used to extract local temporal features through the sequence optimization module and compare them with the real signals, outputting the probability that the signal is a real signal. The mutual information regularization module is used to maximize the latent coding. The mutual information constraint generation process between the generator and the generated signal.

2. The GNSS interference detection system based on generative adversarial networks according to claim 1, characterized in that, The generator is used to segment the input noise sequence into fixed-length segments, merge each noise sequence with the category vector to form a composite input subsequence, and send it to the sequence optimization module. The GRU unit performs segmented feature extraction to generate the corresponding hidden state sequence, and converges it into the hidden state of the last segment as the initial input for parallel multi-step prediction. The parallel mechanism generates the output of multiple subsequences simultaneously. After dimensionality transformation and splicing, a pseudo-signal sample sequence with the same size as the real sample is formed.

3. The GNSS interference detection system based on generative adversarial networks according to claim 2, characterized in that, The parallel multi-step prediction introduces position coding, which is constructed by concatenating relative position coding and channel position coding.

4. The GNSS interference detection system based on generative adversarial networks according to claim 1, characterized in that, The discriminator divides the complete long time-series signal into multiple fixed-length segments and inputs them into the sequence optimization module. The GRU unit extracts the temporal features of each segment, generates the corresponding hidden state sequence, and summarizes them into a feature vector. Based on this feature vector, a classification head performs single-step detection and outputs the probability that the signal is the real signal.

5. The GNSS interference detection system based on generative adversarial networks according to claim 2 or 4, characterized in that, The hidden state of the previous segment is passed to the next segment as context information.

6. The GNSS interference detection system based on generative adversarial networks according to claim 1, characterized in that, The mutual information regularization module includes an auxiliary network. The auxiliary network shares the feature extraction layer with the discriminator and outputs the conditional distribution. During training, the latent encoding is maximized through variational lower bound estimation. Mutual information with the generated samples.

7. The GNSS interference detection system based on generative adversarial networks according to claim 1, characterized in that, It also includes an interference determination module, which dynamically adjusts the positioning strategy based on the probability value output by the discriminator.

8. The GNSS interference detection system based on generative adversarial networks according to claim 1, characterized in that, It also includes a data preprocessing module for performing in-phase component processing on the input signal. Orthogonal components The extraction and normalization process is calculated using the following formula: , The original signal is then normalized and scaled to the [0,1] interval using minimum-maximum normalization.

9. A GNSS interference detection method based on generative adversarial networks, characterized in that, Includes the following steps: Step 1: Collect training data and preprocess it; Step 2: Train the generative adversarial network module in the GNSS interference detection system based on generative adversarial network. The generator generates simulated interference signals based on random noise and latent coding. The preprocessed real samples and simulated interference signals are input together into the discriminator to carry out adversarial training until the training is completed. Step 3: During real-time detection, the data is preprocessed and then input into the trained discriminator. The discriminator outputs the interference probability value to the interference determination module, which dynamically adjusts the positioning strategy based on the probability value output by the discriminator.

Images