A train satellite positioning unsupervised deception detection method based on zero deception samples

Abstract

The application provides a train satellite positioning unsupervised spoofing detection method based on zero spoofing samples. The method comprises the following steps: extracting satellite signal related statistical features, constructing a satellite signal feature time sequence graph in time dimension, and generating a no-label "zero spoofing" training feature sample set, a no-label "zero spoofing" verification feature sample set and a test feature sample set by using the satellite signal feature time sequence graph; training a diffusion reconstruction network by using the no-label "zero spoofing" training feature sample set, verifying the diffusion reconstruction network by using the verification feature sample set and the test feature sample set, and obtaining the trained diffusion reconstruction network; inputting a signal feature time sequence graph of a train in a running process into the diffusion reconstruction network, calculating a reconstruction error index value, and judging whether the train is subjected to a spoofing attack based on the size of the reconstruction error index value. The application realizes real-time detection of spoofing signals in the train running process, and provides effective support for interference protection of a train positioning system.

Description

Technical Field

[0001] This invention relates to the field of train satellite positioning technology, and in particular to an unsupervised deception detection method for train satellite positioning based on zero-deception samples. Background Technology

[0002] Based on the process from satellite signal generation to navigation and positioning, spoofing detection techniques can be categorized into four types: spoofing detection based on navigation data, spoofing detection based on spatial processing, spoofing detection based on radio frequency front-end, spoofing detection based on baseband digital signal processing, spoofing detection based on positioning and navigation results, and spoofing detection based on machine learning. Spoofing detection based on navigation data primarily involves encrypting and authenticating navigation signals to prevent spoofing attacks from altering navigation information at the source. Spoofing detection based on spatial processing mainly involves detecting the signal's angle of arrival and direction of arrival. Spoofing detection based on baseband digital signals primarily involves identifying differences in received signal strength, signal quality monitoring, Doppler shift consistency, and signal arrival time. Detection methods based on positioning and navigation results typically require the introduction of auxiliary sensors resistant to electromagnetic interference, comparing the measured data with the results processed by the receiver terminal. Spoofing detection based on machine learning algorithms primarily involves learning the differences between input features to distinguish between spoofed and genuine signals, thus achieving spoofing interference detection.

[0003] Existing technologies have proposed a sensor fusion method based on particle filters for detecting satellite positioning spoofing interference. This method enhances the robustness of the fusion of satellite navigation systems and inertial navigation systems by utilizing the output data of the IMU (Inertial Measurement Unit), vehicle route information, and road features (such as lanes and edges). This method detects anomalous changes in pseudorange measurements under spoofing attack conditions.

[0004] Existing technologies and solutions have proposed a multi-parameter spoofing interference signal detection algorithm based on a gating mechanism and a parallel CNN-Transformer neural network. This algorithm considers the impact of spoofing interference on various parameters of the receiver during the tracking stage and extracts multiple feature information to form a multi-dimensional time series. After preprocessing the series, the PCTN network is used to extract the temporal relationship and feature information in multiple dimensions of the series. The feature information is then fed into the network for processing to achieve spoofing interference detection. Experimental results show that this method can effectively improve the versatility and generalization of spoofing interference detection algorithms in multiple scenarios.

[0005] The disadvantages of the existing deception interference detection methods mentioned above include:

[0006] Existing methods for detecting deception and interference typically require additional hardware, such as high-precision sensors or extra antennas, to achieve satisfactory detection results. This not only increases the complexity of system implementation but also incurs additional costs, making it difficult to widely apply these technologies in the high-security railway sector.

[0007] Most existing deep learning-based spoofing interference detection methods rely on training with a large number of labeled spoofing interference samples. However, in the railway field, real spoofing interference data is relatively scarce. This leads to problems such as insufficient generalization ability, poor reliability, and low accuracy when these methods are applied to railway scenarios, making it difficult to meet the actual needs of train positioning. Summary of the Invention

[0008] This invention provides an unsupervised spoofing detection method for train satellite positioning based on zero spoofing samples, so as to effectively improve the reliability of satellite positioning such as Beidou.

[0009] To achieve the above objectives, the present invention adopts the following technical solution.

[0010] An unsupervised spoofing detection method for train BeiDou positioning based on zero-spoofing samples includes:

[0011] Digital intermediate frequency signals from train satellite positioning were collected to construct a "zero-spoof sample" training dataset and a "zero-spoof sample" verification dataset that do not contain any spoofing interference. By injecting spoofing signals, a test feature dataset containing spoofing interference was generated.

[0012] Extract relevant statistical features of satellite signals, construct a time series map of satellite signal features in the time dimension, and use the time series map of satellite signal features to generate an unlabeled “zero spoofing” training feature sample set, an unlabeled “zero spoofing” verification feature sample set, and a test feature sample set;

[0013] A diffusion reconstruction network suitable for unsupervised deception interference detection in the field of rail transit is constructed. The diffusion reconstruction network is trained using the unlabeled "zero deception" training feature sample set, and validated using the validation feature sample set and test feature sample set. The reconstruction error of the diffusion reconstruction network is quantified by mean square error and structural similarity index, and the trained diffusion reconstruction network is obtained.

[0014] During the train's operation, it receives satellite digital intermediate frequency signals in real time, extracts relevant statistical features of the satellite signals and generates a signal feature time series diagram, and inputs the signal feature time series diagram into a trained diffusion reconstruction network. The diffusion reconstruction network outputs a model reconstruction diagram, calculates the reconstruction error index value, and determines whether the train has been subjected to a deception attack based on the magnitude of the reconstruction error index value.

[0015] Preferably, the digital intermediate frequency signal for train satellite positioning is collected to construct a "zero-spoof sample" training dataset and a "zero-spoof sample" verification dataset that do not contain any spoofing interference. A test feature dataset containing spoofing interference is then generated by injecting spoofing signals, including:

[0016] During the operation of a train on a specific line, positioning data is collected using a train position reference system to generate a train trajectory file. A satellite navigation signal simulation and acquisition environment is built, and the train trajectory file is injected into the simulation system to reproduce the train operation scenario. Satellite positioning digital intermediate frequency signals of a certain duration are collected to generate a "zero-deception sample" training dataset and a "zero-deception sample" verification dataset that do not contain any deception interference.

[0017] During the reproduction of the train operation scenario, deception interference signals are injected into the “zero-deception sample” training dataset and the “zero-deception sample” verification dataset, which do not contain any deception interference, according to a predetermined time window. Satellite positioning digital intermediate frequency signals for the corresponding time period are collected to generate a test feature dataset containing deception interference.

[0018] Preferably, the step of extracting relevant statistical features of satellite signals, constructing a time-series map of satellite signal features with time as the dimension, and generating an unlabeled "zero-spoofing" training feature sample set using the time-series map of satellite signal features includes:

[0019] The acquired digital intermediate frequency (IF) signal of the train's satellite positioning is sent to a software receiver for processing. Based on the acquisition and tracking loop processing flow, the signal power characteristics, signal correlation peak characteristics, and signal tracking stability characteristics of the satellite positioning IF signal are extracted at frequency f1.

[0020] (1) The carrier noise power spectral density C / N0 is used as the signal power characteristic, which reflects the ratio of the pre-demodulation signal power C to the noise power spectral density N0:

[0021]

[0022] Among them, P C P represents the power of the carrier signal. N It is the noise power within the passband;

[0023] (2) Signal correlation peak characteristics include Delta index, Ratio index, ELP index and Q channel index;

[0024] The formula for calculating the Delta index is:

[0025]

[0026] Among them, IE (t),I L (t),I P (t) represent the in-phase component outputs of the early code, late code, and instant code correlators, respectively;

[0027] The formula for calculating the Ratio index is:

[0028]

[0029] The formula for calculating the ELP metric is:

[0030]

[0031] Among them, Q E (t),Q L (t) represents the quadrature component outputs of the early code and late code correlators, respectively;

[0032] The formula for calculating the Q channel indicator is:

[0033]

[0034] m SQM The indicator identifies the injection of deceptive signals by measuring abnormal energy in the positive traffic channel. If the energy in the positive traffic channel is abnormally high, it indicates that the signal modulation structure or phase consistency has been disrupted.

[0035] Based on a fixed window length, the moving mean and moving variance are calculated for different types of features. The formulas for calculating the moving mean and moving variance are as follows:

[0036]

[0037] Where, μ x (n) represents the mean of feature x at the nth sliding time step. Let represent the variance of feature x at the nth sliding time step, where n represents the number of sliding windows, L represents the size of the sliding windows, and x(i) represents the feature value at time i. The mean values ​​of the Delta, Ratio, ELP, and Q-channel indices at the nth sliding time step are respectively represented by μ. Delta (n), μ Ratio (n), μ ELP (n) and The variance at the nth sliding time step is expressed as follows: and

[0038] (3) Signal tracking stability characteristics include code tracking loop and carrier tracking loop. The output of the code loop discriminator represents the code phase tracking error, which is calculated by the following formula:

[0039]

[0040] Among them, I E and I L Q represents the in-phase component output of the early code and late code correlators, respectively. E and Q L These represent the quadrature component outputs of the early code and late code correlators, respectively;

[0041] The output of the carrier loop discriminator represents the carrier phase tracking error, which is calculated using the following formula:

[0042]

[0043] Among them, I P and Q P These represent the in-phase and quadrature component outputs of the instantaneous code correlator, respectively.

[0044] The mean values ​​of code phase tracking error and carrier phase tracking error at the nth sliding time step are respectively expressed as μ dllDiscr (n) and μ pllDiscr (n), and the variance is expressed as... and The extracted signal power features, signal correlation peak features, and signal tracking stability features comprise 19 specific feature indices. The sampling frequency of the instantaneous features is set to be consistent with the sliding window update interval. The signal statistical feature vector fv at time step t is represented as:

[0045]

[0046] The extracted features are aligned according to the time dimension, and a feature time sequence graph is constructed based on the preset deception detection frequency f3. The feature time sequence graph F at time T is... T Represented as:

[0047]

[0048] The size of the satellite signal feature time series map is (f2 / f3)×19, where (f2 / f3) represents the number of time points contained in the feature time series map and 19 represents the feature dimension. The satellite signal feature time series map is used to generate an unlabeled “zero-spoofing” training feature sample set.

[0049] Preferably, the construction of the diffusion reconstruction network for unsupervised deception interference detection tasks applicable to the rail transit field, using the unlabeled "zero-deception" training feature sample set to train the diffusion reconstruction network, includes:

[0050] Based on the time-series-feature two-dimensional structure of satellite signal feature time series map, the noise prediction network in the diffusion model structure was optimized and improved. The residual network ResNet module was introduced into the encoder and decoder parts of U-Net to replace the convolutional block. The channel attention ECA module was introduced and a one-dimensional temporal convolution downsampling / upsampling module was designed and embedded into the encoder and decoder respectively for temporal feature learning. The downsampling module of the encoder and the corresponding upsampling module in the decoder are connected by skip connections. A diffusion reconstruction network suitable for unsupervised deception interference detection tasks in the field of rail transit was constructed.

[0051] The diffusion reconstruction network is trained using the unlabeled "zero-spoofing" training feature sample set. The training process includes the following steps:

[0052] (1) Initialize training data:

[0053] The unlabeled "zero-deception" training feature sample set is standardized.

[0054] (2) Define the key hyperparameters for model training:

[0055] This includes the maximum number of time steps T in the diffusion process, the noise scheduling function, the learning rate, the optimizer type, the number of training epochs, and the batch size;

[0056] The noise scheduling function is set to linear scaling, expressed as:

[0057]

[0058] in, express and T numbers are generated uniformly between them, β t β represents the noise intensity at time step t. start and β end These are the initial and final values ​​of the noise intensity, respectively.

[0059] (3) The training process of the diffusion reconstruction network is a Markov modeling process with denoising as its goal. The training process includes: a forward diffusion process and a reverse denoising process. After training, the diffusion reconstruction network reconstructs and generates the feature time series map based on the backsampling strategy.

[0060]

[0061] Where, x t This represents the noisy feature time series plot at time step t, ε θ (x t ,t) represents the noise component predicted by the model, σ tx represents the noise amplitude sampled at time step t. t-1 This represents the feature time series diagram of the previous time step after denoising. This process is performed iteratively across the entire time domain from t = T → 0, reconstructing a "zero-deception" feature time series diagram.

[0062] Preferably, the step of using mean squared error and structural similarity indices to quantify the reconstruction error of the diffusion reconstruction network to obtain a trained diffusion reconstruction network includes:

[0063] Calculate the original feature time series map x0 and the reconstructed feature time series map respectively. The reconstruction error between the two components includes MSE and SSIM. The mathematical expression for MSE is:

[0064]

[0065] Where D represents the total number of pixels in the sample. This represents the value of the j-th pixel in the original feature time series map. This represents the value of the j-th pixel in the reconstructed feature time sequence map;

[0066] The mathematical expression for SSIM is:

[0067]

[0068] in, and It is image x0 and average brightness and It is image x0 and variance It is image x0 and The covariance, C1 and C2 are constants used for stability calculations, C1 = (K1L) 2 And C2=(K2L) 2 , where L is the dynamic range of the pixel value;

[0069] The reconstruction error of the i-th sample in the unlabeled "zero-deception" verification feature sample set is expressed as: The mean μ of the reconstruction error of the entire unlabeled “zero-deception” verification feature sample set val and standard deviation σ val They are respectively:

[0070]

[0071] Where, N val This indicates the number of "zero deception" verification feature samples;

[0072] According to the reconstruction error table of the i-th sample in the unlabeled "zero-deception" verification feature sample set. The calculation process calculates the reconstruction error of the j-th test feature sample in the test feature dataset, expressed as: The corresponding Z-score is:

[0073]

[0074] Among them, Z (j) The Z-score reflects the j-th test feature sample;

[0075] Using Z-score as the anomaly detection scoring metric, and combining it with the true labels of samples in the test feature dataset to construct an ROC curve, the false positive rate (FPR) and true positive rate (TPR) are calculated by iterating through different Z-score thresholds.

[0076]

[0077] The area under the ROC curve (AUC) can be obtained through integration or numerical calculation.

[0078]

[0079] When the difference between the AUC value and the threshold 1 is less than the set numerical range, the training process of the diffusion reconstruction network is considered to be over, and the trained diffusion reconstruction network is obtained.

[0080] Preferably, during the train's operation, it receives satellite digital intermediate frequency signals in real time, extracts relevant statistical features of the satellite signals and generates a signal feature time series diagram, inputs the signal feature time series diagram into a trained diffusion reconstruction network, the diffusion reconstruction network outputs a model reconstruction diagram, calculates a reconstruction error index value, and determines whether the train has been subjected to a spoofing attack based on the magnitude of the reconstruction error index value, including:

[0081] During the train's journey, it receives satellite digital intermediate frequency signals in real time, extracts relevant statistical features of the signals and generates a satellite signal feature time series diagram, calls the trained diffusion model and obtains the model reconstruction diagram, calculates reconstruction error indicators such as mean square error and structural similarity, and judges whether the train has been deceived based on the magnitude of the reconstruction error indicator values.

[0082] During the train's journey, it receives satellite digital intermediate frequency (IF) signals in real time. These signals are then sent to a software receiver for processing. The receiver extracts the signal power characteristics, peak value characteristics, and tracking stability characteristics of the satellite IF signals, yielding the signal characteristic time-series diagram F at time T. T , represented as:

[0083] The signal characteristic time sequence diagram F T The input is fed into the trained diffusion reconstruction network, and the reconstruction error index, including mean squared error and structural similarity, is calculated. Using a "zero-deception" verification feature sample set as a benchmark, the Z-score index of the current reconstruction error is calculated. The Z-score value is then compared with a pre-selected threshold τ. * When the Z score is greater than the threshold τ, a comparison is made. * If the Z score is not greater than the optimal threshold τ, then it is determined that the train is under a deception attack at the current moment; * If the time is right, it is determined that the train has not been subjected to a deception attack at the current moment.

[0084] As can be seen from the technical solutions provided by the embodiments of the present invention described above, the method of the present invention can learn and reconstruct the characteristic time sequence map of "zero-spoofing" navigation signals, while generating differentiated responses to the characteristic time sequence map of spoofing satellite signals, thereby achieving real-time detection of spoofing signals during train operation. This method provides effective support for interference protection of train positioning systems, thereby ensuring the reliability of train BeiDou positioning.

[0085] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0086] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0087] Figure 1 A flowchart of an unsupervised deception detection method for train BeiDou positioning based on zero deception samples provided in an embodiment of the present invention;

[0088] Figure 2 This is a schematic diagram illustrating data processing using a sliding window mechanism, provided as an embodiment of the present invention.

[0089] Figure 3 This is a model architecture diagram for unsupervised deception interference detection of zero-deception samples based on a diffusion model, provided in an embodiment of the present invention.

[0090] Figure 4 This invention provides a noise prediction network structure diagram for detecting spoofing interference in BeiDou satellite positioning.

[0091] Figure 5 ROC curves of deception interference test results based on different reconstruction error indices provided in embodiments of the present invention;

[0092] Figure 6 This is a comparison chart of deception interference test performance based on different reconstruction error indices, provided for embodiments of the present invention. Detailed Implementation

[0093] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0094] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0095] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0096] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0097] The application scenario of the unsupervised deception detection method for train satellite positioning based on zero-deception samples provided by this invention is as follows: A BeiDou satellite navigation system intermediate frequency signal acquisition device is installed on the train to collect the digital intermediate frequency signal for train satellite positioning. Key indicators such as signal power, signal correlation peak value, and signal tracking stability are extracted from the signal acquisition and tracking loop to construct a satellite signal feature time series map. Then, the processed feature time series map is input into a diffusion model trained on an unlabeled "zero-deception" training feature sample set to generate a reconstructed feature time series map. The reconstruction error values, such as MSE (Mean Squared Error) and SSIM (Structural Similarity), between the input feature time series map and the reconstructed feature time series map are calculated. Further analysis is conducted to determine the degree of deviation between the reconstruction error of the test samples and the reconstruction error distribution of normal samples, thereby determining whether there are feature anomalies in the test samples. This enables the detection and identification of deception interference signals during train operation.

[0098] The processing flow of an unsupervised spoofing detection method for train satellite positioning based on zero spoofing samples provided in this embodiment of the invention is as follows: Figure 1 As shown, it includes the following steps:

[0099] Step S1: Collect the digital intermediate frequency signal of train satellite positioning, construct a "zero-spoof sample" training dataset and a "zero-spoof sample" verification dataset that do not contain any spoofing interference, and generate a test feature dataset for detection and verification by injecting spoofing signals.

[0100] During the train's operation on the Qinghai-Tibet Railway (Lhasa-Shigatse), a high-precision train position reference system was used to collect train positioning data and continuously generate a 30-minute train trajectory file, which serves as the basis for subsequent experimental testing and verification.

[0101] A satellite navigation signal simulation and acquisition environment was built, and a 30-minute train trajectory file was injected into the simulation system to reproduce the train operation scenario. A 30-minute digital intermediate frequency signal was acquired using the BeiDou satellite navigation system's intermediate frequency signal acquisition device to construct a "zero-spoofing sample" training dataset and a "zero-spoofing sample" verification dataset, both free from any spoofing interference. The acquisition time for the training dataset was 20 minutes, and the acquisition time for the verification dataset was 10 minutes.

[0102] During the scenario reproduction process described above, a test feature dataset for spoofing interference was simultaneously constructed. Five minutes into the scenario, a spoofing interference signal was injected into channel PRN=30. The injected interference type was pseudorange spoofing, increasing at a ramp rate of 1 m / s with a power increase rate of 1 dB / s, and the injection duration was 15 minutes. Satellite digital intermediate frequency (IF) signals for this period were acquired using an IF signal acquisition device for a duration of 20 minutes (including 5 minutes of non-spoofing data and 15 minutes of spoofing data). The resulting test feature dataset containing spoofing interference was used for subsequent model performance testing and verification.

[0103] Step S2: Extract relevant statistical features of satellite signals, construct a time series diagram of satellite signal features with time as the dimension, and form an unlabeled feature sample set corresponding to three different datasets.

[0104] The acquired satellite positioning digital intermediate frequency signal of the train is sent to a software receiver for processing. Based on the acquisition and tracking loop processing flow, three types of statistical features are extracted at frequency f1. In this embodiment, f1 = 1000Hz. Specifically:

[0105] (1) Signal power characteristics

[0106] The carrier-to-noise power spectral density (C / N0) is used as a signal power characteristic. This characteristic reflects the ratio between the pre-demodulation signal power C and the noise power spectral density N0, and is usually expressed in logarithmic form, i.e.:

[0107]

[0108] Among them, P C P represents the power of the carrier signal. N It is the noise power within the passband.

[0109] Typically, spoofing jamming devices transmit signals with higher power than the actual satellite signals. When a GPS receiver is attacked by a spoofing signal, the C / N0 ratio changes. Therefore, C / N0 can be used as a characteristic quantity for spoofing jamming detection. The carrier noise power spectral density of the receiver at a given time t is expressed as (C / N0). t .

[0110] (2) Characteristics of signal correlation peaks

[0111] Signal correlation peak features are primarily extracted using Signal Quality Monitoring (SQM) technology to describe the morphological changes of satellite signal correlation functions under deception interference. When a receiver is attacked by deception interference, the deception signal and the real satellite signal will superimpose and interfere, causing distortion of the correlation peak in the tracking phase. This distortion can be identified by analyzing the correlator output values. Specifically, by acquiring the output values ​​of the Early, Prompt, and Late correlators in the tracking loop and calculating their changes within a specific time window, distortion of the correlation function can be detected, thereby identifying potential deception attacks. Specific metrics include:

[0112] ① Delta indicator

[0113] The Delta index measures the power difference between early and late codes, reflecting the symmetry change of the correlation function. Its calculation formula is:

[0114]

[0115] Among them, I E (t),I L (t),I P (t) represents the in-phase component output of the Early, Late, and Prompt correlators, respectively.

[0116] In the absence of deception interference, the correlation function should be approximately symmetrical, with the difference between the early and late code outputs close to zero; however, deception interference causes a peak shift, resulting in a significant deviation of the value, which can then be used for deception detection.

[0117] ②Ratio index

[0118] The ratio index measures the difference between the instant code and the average of the early / late codes, reflecting the sharpness or distortion of the correlation peak. Its calculation formula can be expressed as:

[0119]

[0120] Without deception interference, the correlation function exhibits a significant peak at the instantaneous code position, with relatively low outputs for the early and late codes, displaying a typical sharp correlation pattern. However, in the presence of deception interference, due to the superposition effect of the interference and real signals, the correlation function peak may be broadened or even show a "flat-top" region, leading to increased outputs for the early and late codes. This, in turn, causes a change in the Ratio index, reflecting a decrease in the sharpness of the correlation peak. Therefore, abnormal changes in the Ratio can serve as a sensitive identification feature for deception attacks.

[0121] ③ ELP (Early-Late Phase) index

[0122] ELP is a signal quality monitoring (SQM) metric built upon carrier phase information, used to detect the phase consistency of the BDS received signal within the correlation structure. This metric extracts the instantaneous phase difference by performing an arctangent operation on the ratio of the in-phase (I) to the quadrature (Q) components output by the early and late code correlators, thereby identifying whether the correlation peak is disturbed by a spoofing signal. Its definition formula is:

[0123]

[0124] Among them, Q E (t),Q L (t) represents the quadrature component outputs of the Early and Late correlators, respectively.

[0125] Under normal, interference-free reception conditions, the complex vectors corresponding to the outputs of the early and late code correlators should have a symmetrical structure, and the ELP index should remain at a stable level. When the BDS signal is subjected to a spoofing attack, the superposition effect between the spoofing signal and the real signal disrupts the original symmetry, causing a significant phase shift between the early and late codes, thus resulting in a significant deviation of the ELP value from the normal range. Therefore, ELP can be widely used in spoofing detection tasks as one of the core features for measuring signal structural consistency.

[0126] ④Q-channel indicator

[0127] The energy of satellite navigation signals is mainly concentrated in the in-phase channel (I-channel), while the output of the quadrature channel (Q-channel) is mostly noise. When spoofing interference exists, the spoofing signal has a carrier phase difference with the real signal, and abnormal energy is generated in the quadrature components. This abnormal energy can be used to detect spoofing attacks. Based on this, the SQM index of the Q-channel is defined as:

[0128]

[0129] m SQM The indicator identifies the injection of deceptive signals by measuring abnormal energy in the positive traffic channel. Abnormally high energy in the positive traffic channel indicates a disruption of the signal modulation structure or phase consistency. Compared to traditional I-channel-based Delta and Ratio indicators, m... SQM It is more sensitive to weak phase perturbations, and therefore uses them as a detection feature for deception interference.

[0130] SQM methods based on instantaneous features (such as Delta, Ratio, ELP, etc.) rely solely on the correlator output value at a single moment for judgment. These methods are susceptible to instantaneous noise interference when facing progressive or slowly varying spoofing attacks, and struggle to effectively capture the changing trends of signal features over time. To improve the sensitivity and robustness of spoofing signal detection, a sliding statistics mechanism is introduced. By calculating the mean and variance of feature values ​​within a sliding window, the fluctuation characteristics and evolutionary trends of time-series features are extracted. Figure 2 This is a schematic diagram illustrating data processing using a sliding window mechanism, as provided in an embodiment of the present invention. Figure 2 As shown, based on a fixed window length, the moving mean and moving variance are calculated for different types of features to enhance the detection capability against weak interference or progressively evolving attacks. The formulas for calculating the moving mean and moving variance are as follows:

[0131]

[0132] Where, μ x (n) represents the mean of feature x at the nth sliding time step. Let represent the variance of feature x at the nth sliding time step, where n represents the number of sliding windows, L represents the size of the sliding window, and x(i) represents the feature value at time i. The mean values ​​of the Delta, Ratio, ELP, and Q-channel indices at the nth sliding time step are respectively represented by μ. Delta (n), μ Ratio (n), μ ELP (n) and The variance at the nth sliding time step is expressed as follows: and In this embodiment, L = 100, which is equivalent to each sliding time step containing a time length of 100ms.

[0133] (3) Signal tracking stability characteristics

[0134] In the software receiver tracking loop processing, the code tracking loop (DLL, Delay Lock Loop) and the carrier tracking loop (PLL, Phase Lock Loop) are key components for achieving continuous signal tracking. Both continuously calibrate the locally generated pseudo-code sequence and carrier signal through a feedback control mechanism, ensuring synchronization with the received satellite signal. The discriminator, as the core module of the loop, directly reflects the tracking error between the code phase and carrier phase, thus quantifying the receiver's signal locking stability. The output of the code loop discriminator represents the code phase tracking error, which can be calculated using the following formula:

[0135]

[0136] Among them, I E and I L Q represents the in-phase component output of the early code and late code correlators, respectively. E and Q L These represent the quadrature component outputs of the early code and late code correlators, respectively.

[0137] The output of the carrier loop discriminator represents the carrier phase tracking error, which can be calculated using the following formula:

[0138]

[0139] Among them, Q P and I P These represent the in-phase and quadrature component outputs of the instantaneous code correlator, respectively.

[0140] When a receiver is subjected to spoofing interference, external false signals will disrupt the receiver's ability to lock onto the real signal, damaging the dynamic balance between the DLL and PLL, manifesting as significant fluctuations or trend deviations in the outputs of the two types of discriminators. Therefore, the outputs of the two discriminators can be used as effective stability features for spoofing interference detection. Simultaneously, to enhance the detection's sensitivity to slowly varying interference and robustness against transient noise, a sliding statistical mechanism is introduced to calculate its mean and variance within the sliding window. The mean values ​​of the code phase tracking error and carrier phase tracking error at the nth sliding time step are respectively expressed as μ. dllDiscr (n) and μ pllDiscr (n), and the variance is expressed as... and

[0141] Ultimately, the extracted statistical features—namely, signal power features, signal correlation peak features, and signal tracking stability features—contain a total of 19 specific feature indices. To achieve unified processing of instantaneous features and sliding window statistical features in the time series modeling process, the sampling frequency of the instantaneous features was set to be consistent with the sliding window update interval. This ensures that different features can be aligned and fused at a unified time scale. In this embodiment, f2 = 10Hz is calculated. Based on this, the signal statistical feature vector fv at time step t is expressed as:

[0142]

[0143] The extracted features are aligned according to the time dimension, and a feature time series graph is constructed based on the preset deception detection frequency f3 (usually f2 > f3). The feature time series graph F at time T is... T Represented as:

[0144]

[0145] The size of the feature time series map is (f2 / f3)×19, where (f2 / f3) represents the number of time points included in the feature time series map, and 19 represents the feature dimension. In this embodiment, f3 = 1Hz, and the feature time series map F can be obtained. T The size is 10×19.

[0146] Step S3: For the BeiDou navigation deception interference detection task in the rail transit train scenario, an unsupervised deception interference detection method based on an improved diffusion model is designed. The diffusion model is trained using an unlabeled "zero deception" training feature sample set containing only normal samples, so that the model can learn and reconstruct the feature time sequence map of normal navigation signals.

[0147] Based on the characteristic evolution law of spoofing interference signals in satellite positioning of rail transit trains, and combined with the time-series-feature two-dimensional structure of the feature time series map constructed in step S2, the diffusion model structure is adapted and improved. A diffusion reconstruction network suitable for unsupervised spoofing interference detection tasks in the rail transit field is designed, enabling it to accurately learn the distribution of "zero-spoofing" feature time series maps and achieve high-fidelity reconstruction, suppressing its ability to fit spoofing samples, and thus generating differentiated responses to abnormal feature time series maps. The model architecture diagram of unsupervised spoofing interference detection based on a diffusion model based on zero-spoofing samples provided in this embodiment of the invention is shown below. Figure 3 As shown. In this embodiment, the noise prediction network in the diffusion model was optimized and improved based on the characteristics of the extracted satellite signal feature time series map.

[0148] The present invention provides a noise prediction network structure diagram for detecting spoofing interference in BeiDou satellite positioning, as shown in the embodiment of the invention. Figure 4As shown, this example uses the U-Net architecture and introduces several key modules. First, considering the potential information loss or insufficient expressive power of traditional convolutional layers when processing complex temporal data, this example introduces ResNet modules in the encoder and decoder parts of U-Net to replace traditional convolutional blocks. ResNet effectively alleviates the gradient vanishing problem in deep networks through residual connections, enhancing the information transfer capability in deep networks, thereby significantly improving the network's feature extraction capability and learning depth. Second, to further improve the network's sensitivity to key features, an Efficient Channel Attention (ECA) module is introduced. This module dynamically adjusts the weights of different channels, enabling the network to adaptively focus on important information when processing complex features, thereby strengthening its ability to learn key features and improving the network's learning efficiency. Furthermore, to better capture the temporal variation characteristics of the feature time-series maps, a one-dimensional temporal convolutional downsampling / upsampling module was designed and embedded into the Encoder and Decoder respectively for temporal feature learning. This design enhances the network's modeling ability in the temporal dimension, thereby improving the accuracy and effectiveness of noise prediction. The downsampling module in the Encoder and the corresponding upsampling module in the Decoder are connected via skip connections to preserve shallow feature information. This mechanism effectively avoids information loss during reconstruction and improves the network's convergence speed and overall performance.

[0149] S3.2. The improved diffusion model is trained using the unlabeled "zero-deception" training feature sample set constructed in step S2. The training process includes the following steps:

[0150] (1) Initialize training data:

[0151] Standardize the training feature sample set to ensure that the input dimensions are consistent and fit the model structure;

[0152] (2) Define the key hyperparameters for model training:

[0153] This includes the maximum number of time steps T in the diffusion process, the noise scheduling function (such as linear / cosine), the learning rate, the optimizer type (such as Adam), the number of training epochs, and the batch size.

[0154] In this embodiment, the maximum number of time steps T is set to 1000, the Adam optimizer is used, and the learning rate is set to 2.5 × 10⁻⁶. -5 The training epochs are set to 300, and the batch size is set to 256. The noise scheduling function is set to linear scaling, as follows:

[0155]

[0156] in, express and T numbers are generated uniformly between them, β t β represents the noise intensity at time step t. start and β end These are the initial and final values ​​of the noise intensity, respectively, and are set to 0.0015 and 0.0195.

[0157] (3) Perform the model training process:

[0158] The training process of a diffusion model is a Markov modeling process aimed at denoising, designed to learn how to recover the original clean samples from noisy data. Its training process mainly consists of two stages:

[0159] ① Forward diffusion process:

[0160] In the forward process, Gaussian noise is added to the "zero-deception" feature time series map x0 sampled from the unlabeled "zero-deception" training feature sample set through a fixed noisy Markov chain, generating a series of noisy feature time series maps x1, x2, ..., x T The noise addition process can be represented as:

[0161]

[0162] in, α represents the cumulative signal retention coefficient for the first t steps. s =1-β s ε represents standard Gaussian noise, x t This represents the time sequence diagram of the noisy features generated at step t.

[0163] ② Reverse denoising process:

[0164] The main purpose of the reverse process is to extract the noisy feature time series map x t The original "zero-deception" feature time series map x0 is recovered. This process is based on modeling the forward diffusion chain and relies on a noise prediction network (usually using a U-Net structure) to fit the noisy feature time series map x0. t The mapping relationship between the noise component ε and the original noise ε enables accurate estimation of the noise component ε in the time series graph of each noisy feature.

[0165] Noise prediction network ε θ The modeling objective is to, at any time step t, analyze the noisy feature time series graph x. t To accurately estimate the noise ε in the network, the most commonly used loss function during network training is the mean squared error loss.

[0166]

[0167] The time step t is typically randomly and uniformly sampled from [1, T]. With minimizing the loss function as the training objective, the network parameters θ are iteratively optimized using the gradient descent algorithm to improve the network's ability to accurately fit noise under different levels of added noise.

[0168] After training, the diffusion model reconstructs and generates the feature time series map based on the backsampling strategy:

[0169]

[0170] Where, x t This represents the noisy feature time series plot at time step t, ε θ (x t ,t) represents the noise component predicted by the model, σ t x represents the noise amplitude sampled at time step t. t-1 This represents the feature time series diagram of the previous time step after denoising. This process is performed iteratively across the entire time domain from t = T → 0, ultimately obtaining the complete "zero-deception" feature time series diagram x0.

[0171] Through the above training, the diffusion model can achieve the extraction of noisy feature time series graphs x with arbitrary noise intensities. t The time series diagram x0 of the "zero deception" feature of the original input is gradually reconstructed.

[0172] (4) Verify model performance:

[0173] The model performance was evaluated using a "zero-deception" validation feature sample set. Reconstruction error evaluation metrics such as mean squared error (MSE) and structural similarity (SSIM) between the input feature time series map and the reconstructed feature time series map were calculated to assess the model's learning ability and generalization effect on normal samples.

[0174] S3.3 During the training process, the model parameters with the minimum loss function value are saved and used as the final detection model parameters for subsequent experimental verification and online detection and identification of deception interference during train operation.

[0175] Step S4: Input the "zero-deception" verification feature sample set and the test feature sample set into the trained diffusion model to obtain the model reconstruction map. Mean squared error and structural similarity indices are used to quantify the model reconstruction error. By analyzing the deviation of the test feature sample reconstruction error from the "zero-deception" verification feature sample reconstruction error distribution, it is determined whether there are anomalies in the test feature samples, thus achieving the detection and identification of deception signals. This includes:

[0176] S4.1. Use the model saved in step S3.3 for testing and verification, and generate a model reconstruction diagram, specifically as follows:

[0177] (1) Initialization of feature time sequence map

[0178] The "zero-spoofing" verification feature sample set and test feature sample set obtained in step S2 are standardized to ensure that the size, number of channels, and timing structure of the input data are consistent with the training model in order to adapt to the network input interface.

[0179] (2) Noise addition processing of feature time sequence maps

[0180] Noise is added to the original feature time series map x0 according to a preset time step to construct a noisy feature time series map x. t The calculation formula is as follows:

[0181]

[0182] in, The cumulative signal retention coefficient for the first t steps is represented by ε, and the standard Gaussian noise is represented by x. t This represents the noisy feature time sequence diagram generated at step t. In this embodiment, 10 time steps are used to obtain the noisy feature time sequence diagram, namely t=[10,110,210,310,410,510,610,710,810,910].

[0183] (3) Predicting noise

[0184] Call a pre-trained noise prediction network (such as U-Net) to add noise to the time sequence image x. t Using the corresponding time step t as input, predict the noise components contained in the current noisy feature time series map:

[0185]

[0186] (4) Gradual noise reduction and reconstruction

[0187] Based on the predicted noise information, the denoising and restoration process is executed step by step according to the reverse sampling path set by the scheduler, ultimately generating a denoised sample estimate.

[0188]

[0189] Where σ t The residual noise intensity at each step, In this embodiment, the PLMS accelerated sampling method is used for denoising and reconstruction. A denoising operation is performed every 10 time steps. The denoising process is accelerated by multi-step prediction and correction, making the generation process more efficient while maintaining high generation quality.

[0190] (5) Reconstruction error calculation

[0191] Calculate the original feature time series map x0 and the reconstructed feature time series map respectively. The reconstruction error between the two is used to measure the model's fitting ability. Typical error metrics include:

[0192] ①Mean Squared Error (MSE):

[0193] Mean squared error (MSE) is a fundamental metric that measures the numerical difference between the reconstructed model result and the original input. It is expressed as the average of the squared errors between the predicted and actual values. Its mathematical expression is:

[0194]

[0195] Where D represents the total number of pixels in the sample. This represents the value of the j-th pixel in the original feature time series map. This represents the value of the j-th pixel in the reconstructed feature time series map. The smaller the MSE value, the closer the model's reconstruction result is to the original input, i.e., the higher the reconstruction accuracy.

[0196] ② Structural Similarity Index (SSIM):

[0197] SSIM is a metric for measuring image quality. It aims to simulate the human visual system's perception of structural similarity, measuring the similarity between two images in terms of structure, brightness, and contrast. It is widely used to evaluate image reconstruction quality. It is defined as:

[0198]

[0199] in, and It is image x0 and average brightness and It is image x0 and variance It is image x0 and The covariance, C1 and C2 are constants used for stability calculations. To prevent the denominator from being zero, they are usually set as C1 = (K1L). 2 And C2=(K2L) 2 , where L is the dynamic range of pixel values. The calculated SSIM range is [0,1], and the SSIM value reaches its maximum value of 1 when two images are identical. The smaller the SSIM value, the worse the reconstruction quality.

[0200] MSE measures the overall error at the pixel level, while SSIM measures the structural fidelity of the image. Used together, they comprehensively reflect the model's reconstruction capability from two dimensions: error magnitude and structural similarity. In this embodiment, both MSE and SSIM reconstruction error metrics were calculated for subsequent detection and recognition.

[0201] Based on this, the reconstruction error of the model for the "zero deception" verification feature samples and test feature samples is obtained.

[0202] S4.2. By analyzing the deviation of the reconstruction error of the test feature samples from the "zero deception" verification feature sample reconstruction error distribution, it is determined whether there are feature anomalies in the test feature samples, thereby realizing the detection and identification of deception interference signals. Specifically:

[0203] (1) Calculate the statistical characteristics of the reconstruction error of the "zero deception" verification feature samples.

[0204] Based on the reconstruction error of the "zero-deception" verification feature sample calculated in step S4.1, the statistic of the reconstruction error of the entire "zero-deception" verification feature sample set is calculated. The reconstruction error of the i-th sample in the "zero-deception" verification feature sample set is expressed as: The mean μ of the reconstruction error of the entire "zero deception" verification feature sample set is... val and standard deviation σ val They are respectively:

[0205]

[0206] Where, N val This represents the number of "zero-deception" verification feature samples. The mean and standard deviation of this value serve as reference standards for calculating the Z-score in subsequent anomaly detection.

[0207] (2) Calculate the Z-score of the reconstruction error of the test feature dataset.

[0208] The mean μ of the feature sample reconstruction error is verified based on "zero deception". val and standard deviation σ val The reconstruction error of the test feature samples is standardized, and the reconstruction error of the j-th sample in the test feature sample set is expressed as: The corresponding Z-score is:

[0209]

[0210] Among them, Z (j)The Z-score reflects the degree of deviation of its reconstruction error from the reconstruction error distribution of the "zero-deception" verification feature sample. A higher Z-score indicates that the sample is more likely to be a deceptively manipulated sample. In this embodiment, the average Z-score calculated at 10 different time steps t = [10, 110, 210, 310, 410, 510, 610, 710, 810, 910] is used for subsequent detection performance index calculations.

[0211] (3) Construct the ROC curve and calculate the AUC value

[0212] Using the Z-score as the anomaly detection metric, and combining it with the true labels of the samples in the test set (0 for normal, 1 for deception), an ROC curve (Receiver Operating Characteristic Curve) is constructed. By iterating through different Z-score thresholds, the false positive rate (FPR) and true positive rate (TPR) at each threshold are calculated.

[0213]

[0214] Finally, the area under the ROC curve (AUC) is obtained through integration or numerical calculation:

[0215]

[0216] The AUC value reflects the ability of the detection model to distinguish between "normal" and "deceptive" samples. The closer the AUC is to 1, the stronger the model's ability to distinguish based on reconstruction error. An AUC close to 0.5 indicates that the model has almost no ability to distinguish between "normal" and "deceptive" samples.

[0217] In this embodiment, we used Z-scores of two error metrics, mean squared error (MSE) and structural similarity (SSIM), to evaluate the performance of the detection model. Figure 5 ROC curves of deception interference test results based on different reconstruction error indices provided in embodiments of the present invention are shown below. Figure 5As shown in Table 1, the ROC curves were plotted based on different error metrics. It can be observed that when the model combines the sum of these two error metrics for detection, its recognition ability is better, with an AUC (area under the curve) value reaching 99.78%. The AUC values ​​of different error metrics in the experiment are shown in Table 1. The experimental results show that combining the MSE and SSIM error metrics for detection improves the AUC value by 3.05% and 0.12% respectively compared to using SSIM or MSE alone, demonstrating the effectiveness of the combined metric method in improving detection accuracy.

[0218] Table 1. AUC values ​​obtained from different reconstruction error indices

[0219] Reconstruction error index AUC SSIM 96.83％ MSE 99.66％ MSE+SSIM 99.78％

[0220] (4) Determine the optimal threshold and calculate accuracy, precision, recall, and F1 score.

[0221] To further quantify the model's performance in classifying test samples based on reconstruction error, the quantile method was used to determine the optimal threshold τ. * Then, the classification performance index is calculated. The specific steps are as follows:

[0222] ① Select a threshold determination method to determine the optimal threshold τ * Common methods include maximizing the Youden exponent and using quantiles. In this embodiment, the quantile method is used to set the threshold. Specifically, the Z-score is calculated based on the reconstruction error of the validation feature sample set, and the 95th quantile of the Z-score distribution is used as the optimal threshold τ. * .

[0223] Based on this, the optimal thresholds for SSIM, MSE, and MSE+SSIM are 1.3341, 0.2849, and 1.4102, respectively.

[0224] ②Based on the optimal threshold τ * Divide the prediction labels (0 for normal, 1 for deception):

[0225]

[0226] in, Let be the predicted label of the j-th sample by the model.

[0227] ③ Calculate the detection performance indicators

[0228] After obtaining the predicted labels, the following indicators can be calculated based on the statistics of the confusion matrix:

[0229] (a) Accuracy: The proportion of the model's overall predictions that are correct.

[0230]

[0231] (b) Precision: The proportion of samples predicted as anomalous that are actually anomalous.

[0232]

[0233] (c) Recall: The proportion of outlier samples that are correctly identified.

[0234]

[0235] (d) F1 Score: The harmonic mean of precision and recall, measuring overall performance.

[0236]

[0237] Wherein, TP represents the number of samples that were actually deceptive but were predicted as deceptive; TN represents the number of samples that were actually normal but were predicted as normal; FP represents the number of samples that were actually normal but were predicted as deceptive; and FN represents the number of samples that were actually deceptive but were predicted as normal.

[0238] In this embodiment, the accuracy, precision, recall, and F1 score based on SSIM, MSE, and MSE+SSIM are shown in Table 2. Figure 6 This is a comparison chart of deception interference test performance based on different reconstruction error indices provided in an embodiment of the present invention.

[0239] Table 2 Detection performance obtained from different reconstruction error indices

[0240] Reconstruction error index accuracy Accuracy Recall rate F1 Score SSIM 91.83％ 97.29％ 91.67％ 94.39％ MSE 96.58％ 96.23％ 99.33％ 97.76％ MSE+SSIM 97.83％ 97.60％ 99.56％ 98.57％

[0241] Experimental results show that using SSIM or MSE alone, or combining both, demonstrates the model's ability to detect deception interference to some extent. However, combining the two error metrics, MSE and SSIM, improves accuracy by 6.53% and 1.29%, precision by 0.32% and 1.42%, recall by 8.61% and 0.23%, and F1 score by 4.43% and 0.83%, respectively, compared to using SSIM or MSE alone. This proves that combining error metrics to identify deception is a more reliable detection scheme, suitable for unsupervised deception interference detection tasks, and can effectively improve detection accuracy.

[0242] Step S5: During the train's operation, the train receives BeiDou satellite digital intermediate frequency signals in real time, extracts relevant statistical features of the signals and constructs a satellite signal feature time series diagram, calls the trained diffusion model and obtains the model reconstruction diagram, calculates reconstruction error indicators such as mean square error and structural similarity, and judges whether the train has been subjected to a deception attack based on the magnitude of the reconstruction error indicator values.

[0243] During the train's journey, it receives real-time digital intermediate frequency (IF) signals from the BeiDou satellite system. The collected satellite positioning IF signals are then sent to a software receiver for processing. Based on the acquisition and tracking loop's processing flow, three types of statistical features are extracted, including signal power, signal peak value, and signal tracking stability. A feature time series diagram is then constructed according to step S2, resulting in the feature time series diagram F for time T. T , represented as:

[0244] The feature time sequence map F T Input the detection model saved in step S3, generate the model reconstruction map according to step S4, and calculate reconstruction error indices such as mean square error and structural similarity; using the "zero deception" verification feature sample set as a benchmark, calculate the Z-score index of the current reconstruction error, and compare the Z-score value with the optimal threshold τ obtained in step S4. * When the Z score is greater than the optimal threshold τ, the comparison is made. * If the Z score is not greater than the optimal threshold τ, then it is determined that the train is under a deception attack at the current moment; * If the system detects a spoofing attack, it determines that the train is not under attack at that moment. This enables unsupervised spoofing detection of train satellite positioning.

[0245] In summary, this invention presents an unsupervised deception detection method for train satellite positioning based on zero-deception samples, used for identifying deception interference signals. This method primarily extracts relevant statistical features of satellite signals, constructs a time-series map of satellite signal features, captures dynamic changes in the signal, and forms an unlabeled feature sample set. An unsupervised deception interference detection model based on an improved diffusion model is designed. The model is trained using the unlabeled "zero-deception" training feature sample set, enabling it to accurately learn the distribution patterns of interference-free samples and achieve high-fidelity reconstruction of normal signals. For abnormal samples, the model can effectively identify their deviation from the "zero-deception" normal samples through differentiated responses, thereby suppressing the fitting ability to deception samples and achieving deception detection based on reconstruction error. This method effectively overcomes the dependence of traditional supervised learning methods on large amounts of labeled data, significantly improving the train satellite positioning accuracy.

[0246] This invention proposes an unsupervised spoofing detection method for train BeiDou positioning based on zero-spoofing samples. This method fully utilizes the digital intermediate frequency signal of BeiDou satellite navigation and positioning, extracting three types of statistical features: signal power, signal correlation peak value, and signal tracking stability. A time-series map of satellite signal features is constructed using time as the dimension. An unlabeled "zero-spoofing" verification feature sample set is used to train a diffusion model, enabling the model to learn and reconstruct the feature time-series map of the "zero-spoofing" navigation signal, while generating differentiated responses to the feature time-series map of spoofing satellite signals. This allows for real-time detection of spoofing signals during train operation. This method provides effective support for interference protection of train positioning systems, thereby ensuring the reliability of train BeiDou positioning.

[0247] This invention is applicable to different types of satellite navigation systems under observable conditions. It is universally applicable to different types of satellite navigation deception and interference modes, various railway line network conditions, different types of surrounding environmental conditions, and different network scales, and has significant engineering application value.

[0248] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0249] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0250] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0251] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An unsupervised spoofing detection method for train satellite positioning based on zero-spoofing samples, characterized in that, include: Digital intermediate frequency signals from train satellite positioning were collected to construct a "zero-spoof sample" training dataset and a "zero-spoof sample" verification dataset that do not contain any spoofing interference. By injecting spoofing signals, a test feature dataset containing spoofing interference was generated. Extract relevant statistical features of satellite signals, construct a time series map of satellite signal features in the time dimension, and use the time series map of satellite signal features to generate an unlabeled "zero spoofing" training feature sample set, an unlabeled "zero spoofing" verification feature sample set, and a test feature sample set; A diffusion reconstruction network suitable for unsupervised deception interference detection in the field of rail transit is constructed. The diffusion reconstruction network is trained using the unlabeled "zero deception" training feature sample set, and validated using the validation feature sample set and test feature sample set. The reconstruction error of the diffusion reconstruction network is quantified by mean square error and structural similarity index, and the trained diffusion reconstruction network is obtained. During the train's operation, it receives satellite digital intermediate frequency signals in real time, extracts relevant statistical features of the satellite signals and generates a signal feature time series diagram, and inputs the signal feature time series diagram into a trained diffusion reconstruction network. The diffusion reconstruction network outputs a model reconstruction diagram, calculates the reconstruction error index value, and determines whether the train has been subjected to a deception attack based on the magnitude of the reconstruction error index value.

2. The method according to claim 1, characterized in that, The aforementioned method involves collecting digital intermediate frequency signals for train satellite positioning, constructing a "zero-spoof sample" training dataset and a "zero-spoof sample" verification dataset free from any spoofing interference, and generating a test feature dataset containing spoofing interference by injecting spoofing signals, including: During the operation of a train on a specific line, positioning data is collected using a train position reference system to generate a train trajectory file. A satellite navigation signal simulation and acquisition environment is built, and the train trajectory file is injected into the simulation system to reproduce the train operation scenario. Satellite positioning digital intermediate frequency signals of a certain duration are collected to generate a "zero-deception sample" training dataset and a "zero-deception sample" verification dataset that do not contain any deception interference. During the reproduction of the train operation scenario, deception interference signals are injected into the "zero deception sample" training dataset and the "zero deception sample" verification dataset, which do not contain any deception interference, according to a predetermined time window. Satellite positioning digital intermediate frequency signals for the corresponding time period are collected to generate a test feature dataset containing deception interference.

3. The method according to claim 1, characterized in that, The extraction of relevant statistical features of satellite signals, the construction of a time-series map of satellite signal features, and the generation of an unlabeled "zero-spoof" training feature sample set using the time-series map of satellite signal features include: The acquired satellite positioning digital intermediate frequency signal of the train is sent to the software receiver for processing. Following the processing flow of the acquisition and tracking loop, the signal is processed at a specific frequency. Extract the signal power characteristics, signal correlation peak characteristics, and signal tracking stability characteristics of the satellite positioning digital intermediate frequency signal: (1) Using carrier noise power spectral density As a signal power characteristic, this signal power characteristic reflects the pre-demodulation signal power. With noise power spectral density The ratio between them: （1） in, Indicates the power of the carrier signal. It is the noise power within the passband; (2) Signal correlation peak characteristics include Delta index, Ratio index, ELP index and Q channel index; The formula for calculating the Delta index is: （2） in, These represent the in-phase component outputs of the early code, late code, and instant code correlators, respectively. The formula for calculating the Ratio index is: （3） The formula for calculating the ELP metric is: （4） in, These are the quadrature component outputs of the early code and late code correlators, respectively; The formula for calculating the Q channel indicator is: （5） The indicator identifies the injection of deceptive signals by measuring abnormal energy in the positive traffic channel. If the energy in the positive traffic channel is abnormally high, it indicates that the signal modulation structure or phase consistency has been disrupted. Based on a fixed window length, the moving mean and moving variance are calculated for different types of features. The formulas for calculating the moving mean and moving variance are as follows: （6） （7） in, Representation of features In the The mean of each sliding time step, Representation of features In the The variance of each sliding time step Indicates the number of sliding windows. Indicates the size of the sliding window. Indicates the first The eigenvalues ​​at time n, the Delta indicator, the Ratio indicator, the ELP indicator, and the Q-channel indicator at time n. The mean values ​​of the sliding time steps are respectively expressed as: , , and ; in the The variances of each sliding time step are expressed as follows: , , and ; (3) Signal tracking stability characteristics include code tracking loop and carrier tracking loop. The output of the code loop discriminator represents the code phase tracking error, which is calculated by the following formula: （8） in, and These represent the in-phase component outputs of the early and late code correlators, respectively. and These represent the quadrature component outputs of the early code and late code correlators, respectively; The output of the carrier loop discriminator represents the carrier phase tracking error, which is calculated using the following formula: （9） in, and These represent the in-phase and quadrature component outputs of the instantaneous code correlator, respectively. In the The mean values ​​of code phase tracking error and carrier phase tracking error for each sliding time step are respectively expressed as: and The variances are respectively expressed as and The extracted signal power features, signal correlation peak features, and signal tracking stability features comprise 19 specific feature indices. The sampling frequency of the instantaneous features is set to be consistent with the sliding window update interval. In time step Signal statistical feature vector Represented as: The extracted features are aligned according to the time dimension and based on a preset deception detection frequency. Construct a feature time series graph at time 100000. Feature time series diagram Represented as: The size of the satellite signal characteristic time series diagram is: , The number of time points contained in the feature time series plot is represented by 19, which represents the feature dimension. The satellite signal feature time series plot is used to generate an unlabeled "zero-spoof" training feature sample set.

4. The method according to claim 1, characterized in that, The aforementioned construction of a diffusion reconstruction network suitable for unsupervised deception interference detection tasks in the rail transit field, using the unlabeled "zero-deception" training feature sample set to train the diffusion reconstruction network, includes: Based on the time-series-feature two-dimensional structure of satellite signal feature time series map, the noise prediction network in the diffusion model structure was optimized and improved. The residual network ResNet module was introduced into the encoder and decoder parts of U-Net to replace the convolutional block. The channel attention ECA module was introduced and a one-dimensional temporal convolution downsampling / upsampling module was designed and embedded into the encoder and decoder respectively for temporal feature learning. The downsampling module of the encoder and the corresponding upsampling module in the decoder are connected by skip connections. A diffusion reconstruction network suitable for unsupervised deception interference detection tasks in the field of rail transit was constructed. The diffusion reconstruction network is trained using the unlabeled "zero-spoofing" training feature sample set. The training process includes the following steps: (1) Initialize training data: The unlabeled "zero-deception" training feature sample set is standardized. (2) Define the key hyperparameters for model training: Including the maximum number of time steps in the diffusion process Noise scheduling function, learning rate, optimizer type, number of training epochs, and batch size; The noise scheduling function is set to linear scaling, expressed as: （10） in, express and Evenly generated between Number, Indicates the first Noise intensity at each time step and These are the initial and final values ​​of the noise intensity, respectively. (3) The training process of the diffusion reconstruction network is a Markov modeling process with denoising as its goal. The training process includes: a forward diffusion process and a reverse denoising process. After training, the diffusion reconstruction network reconstructs and generates the feature time series map based on the backsampling strategy. （13） in, Indicates time step Noisy feature time series diagram, This represents the noise component predicted by the model. This indicates the time step. The noise amplitude of the sample. This represents the feature time sequence diagram of the previous time step after denoising. This process occurs in... Full-time-domain iterative execution reconstructs the "zero-deception" feature time sequence graph. .

5. The method according to claim 1, characterized in that, The method of using mean squared error and structural similarity metrics to quantify the reconstruction error of the diffusion reconstruction network to obtain a trained diffusion reconstruction network includes: Calculate the original feature time series plots respectively and reconstructed feature time series diagram The reconstruction error between the two components includes MSE and SSIM. The mathematical expression for MSE is: （17） in, This represents the total number of pixels in the sample. Represents the original feature time series diagram. The value of each pixel. Represents the reconstructed feature time sequence diagram. The value of each pixel; The mathematical expression for SSIM is: （18） in, and It is an image and average brightness and It is an image and variance It is an image and covariance, and These are constants used for stability calculations. and ,in It is the dynamic range of pixel values; The first unlabeled "zero deception" verification feature sample set The reconstruction error of each sample is expressed as: The mean of the reconstruction error of the entire unlabeled "zero deception" verification feature sample set and standard deviation They are respectively: （19） in, This indicates the number of "zero deception" verification feature samples; According to the labelless "zero deception" verification feature sample set, the first... Reconstruction error table for each sample The calculation process calculates the first feature in the test feature dataset. The reconstruction error of each test feature sample is expressed as: The corresponding Z-score is: （20） in, Reflecting the Z-score of each test feature sample; Using Z-score as the anomaly detection scoring metric, and combining it with the true labels of samples in the test feature dataset to construct an ROC curve, the false positive rate (FPR) and true positive rate (TPR) are calculated by iterating through different Z-score thresholds. （21） The area under the ROC curve (AUC) can be obtained through integration or numerical calculation. （22） When the difference between the AUC value and the threshold 1 is less than the set numerical range, the training process of the diffusion reconstruction network is considered to be over, and the trained diffusion reconstruction network is obtained.

6. The method according to claim 1, characterized in that, During the train's operation, it receives satellite digital intermediate frequency signals in real time, extracts relevant statistical features of the satellite signals, and generates a signal feature time series diagram. This signal feature time series diagram is then input into a trained diffusion reconstruction network. The diffusion reconstruction network outputs a model reconstruction diagram, calculates a reconstruction error index value, and determines whether the train has been subjected to a spoofing attack based on the magnitude of the reconstruction error index value. This includes: During the train's journey, it receives satellite digital intermediate frequency signals in real time, extracts relevant statistical features of the signals and generates a satellite signal feature time series diagram, calls the trained diffusion model and obtains a model reconstruction diagram, calculates the mean square error and structural similarity reconstruction error index, and determines whether the train has been subjected to a deception attack based on the magnitude of the reconstruction error index value; During its journey, the train receives satellite digital intermediate frequency (IF) signals in real time. These signals are then sent to a software receiver for processing. The receiver extracts the signal power characteristics, peak value characteristics, and tracking stability characteristics of the satellite IF signals to obtain the train's time. Signal characteristic timing diagram , represented as: ; Signal feature time sequence diagram The input is fed into the trained diffusion reconstruction network, and the reconstruction error index, including mean squared error and structural similarity, is calculated. Using a "zero-deception" verification feature sample set as a benchmark, the Z-score of the current reconstruction error is calculated, and the Z-score is compared with a pre-selected threshold. When the Z score is greater than the threshold, a comparison is made. If the Z-score is not greater than the optimal threshold, then it is determined that the train is under a deception attack at the current moment; If the time is right, it is determined that the train has not been subjected to a deception attack at the current moment.

Images