GNSS real-time deception interference detection method and system based on deep learning

Abstract

The present invention discloses a GNSS real-time deception interference detection method and system based on deep learning, which belongs to the field of satellite navigation signal processing. The method comprises the following steps: collecting Doppler frequency shift signals of a GNSS receiver in real time to obtain an initial training data set; constructing an LSTM prediction model with an incremental learning mechanism, comprising a double-layer LSTM network and a fully connected layer, wherein the double-layer LSTM network introduces a gating mechanism, and using the initial training data set to add Gaussian white noise to the prediction model for model training; performing sliding window processing on the initial training data set, and extracting a multidimensional feature vector; inputting the trained prediction model for prediction to obtain a model prediction value; calculating the residual between the model prediction value and the true value, calculating the standard deviation of the residual after excluding a warm-up period, and dynamically setting a threshold based on the standard deviation of the residual; and determining that a signal interference event has occurred when the absolute value of the residual exceeds the threshold and the sampling point processed by the sliding window is within a preset interference time window.

Description

Technical Field

[0001] The present invention belongs to the field of satellite navigation signal processing, and specifically relates to a GNSS real-time deception interference detection method and system based on deep learning. Background Art

[0002] As the spatial and temporal benchmark of the modern information society, the Global Navigation Satellite System (GNSS) has been widely used in key sectors such as transportation, communications networks, financial transactions, and national defense. Over 80% of industries worldwide rely on the high-precision positioning and timing services provided by GNSS. However, GNSS signals have inherent flaws during transmission, such as extremely low landing power, with typical signal strength around -160dBW, far below the ambient noise level (-140dBW). This deficiency makes GNSS signals particularly vulnerable to spoofing and interference.

[0003] Current mainstream interference detection technologies include SQM (Signal Quality Monitoring) and STAP (Space-Time Adaptive Processing). The former is based on single-dimensional indicators such as carrier-to-noise ratio and pseudorange residual, but relies on empirical threshold settings and has a high misjudgment rate in complex electromagnetic environments. The latter requires support from multiple antenna arrays, has high hardware costs, and cannot detect covert deceptive interference.

[0004] Detection methods based on deep learning can greatly improve the shortcomings of traditional detection methods and can jointly analyze signal characteristics from multiple dimensions of time, frequency and space. Although some studies have introduced deep learning into GNSS interference detection, there are still the following key bottlenecks: First, in the long-sequence prediction of traditional deep learning models, the prediction errors gradually accumulate over time, which can easily lead to the failure of subsequent predictions; second, static models have limitations, and fixed model parameters may not be able to adapt to the dynamic changes of signal characteristics. In addition, fixed threshold detection methods have a high false alarm rate in complex environments, making it difficult to distinguish between real interference and noise fluctuations. Summary of the Invention

[0005] The purpose of the present invention is to overcome the problems of long-term prediction attenuation and static threshold sensitivity, and propose a GNSS real-time deception interference detection method and system based on deep learning, which can dynamically update model parameters, adapt to environmental changes, and accurately detect GNSS signal interference.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a method for detecting GNSS real-time spoofing interference based on deep learning, comprising the following steps:

[0008] Collect the Doppler frequency shift signal of the GNSS receiver in real time to obtain the initial training data set;

[0009] An LSTM prediction model with an incremental learning mechanism was constructed. Using the initial training dataset, Gaussian white noise was added to the LSTM prediction model with the incremental learning mechanism for model training. The trained LSTM prediction model with the incremental learning mechanism was obtained. The LSTM prediction model with the incremental learning mechanism included a two-layer LSTM network and a fully connected layer. A gating mechanism was introduced into the two-layer LSTM network.

[0010] After performing sliding window processing on the initial training data set, a multidimensional feature vector is extracted; the multidimensional feature vector is input into the trained LSTM prediction model with an incremental learning mechanism for prediction to obtain the model prediction value;

[0011] Calculate the residual between the model prediction value and the real value of the Doppler frequency shift signal collected in real time, calculate the standard deviation of the residual after excluding the warm-up period, and dynamically set the threshold based on the standard deviation of the residual;

[0012] The absolute value of the residual is calculated. When the absolute value of the residual exceeds a threshold and the sampling point processed by the sliding window is within a preset interference time window, it is determined that a signal interference event has occurred.

[0013] Furthermore, the two-layer LSTM network includes an input gate, a forget gate, a cell state update, an output gate and a hidden state, and the multidimensional feature vector includes an instantaneous value, a sliding mean, a standard deviation, a linear trend coefficient and a fast Fourier transform frequency domain component.

[0014] Furthermore, the multi-dimensional feature vector is input into the trained LSTM prediction model with incremental learning mechanism for prediction. The specific prediction is as follows:

[0015] Part of the historical training data is randomly selected and mixed with the real-time collected Doppler frequency shift signal, and the prediction model parameters are dynamically updated using a hierarchical update strategy with weight freezing.

[0016] Furthermore, the hierarchical update strategy with weight freezing dynamically updates the prediction model parameters, including a gradient norm threshold and a hierarchical weight update amount constraint. The gradient norm threshold and the hierarchical weight update amount constraint include:

[0017] The gradient norm threshold range is set using the gradient clipping method;

[0018] The weight update amount of the two-layer LSTM network is within the first threshold range of the initial weight;

[0019] The weight update amount of the fully connected layer is within the second threshold range of the initial weight.

[0020] Furthermore, the hierarchical update strategy with weight freezing adopts an incremental learning rule, which is specifically:

[0021]

[0022] The constraints of the incremental learning rule are:

[0023]

[0024]

[0025] in, is the updated LSTM layer parameter, is the LSTM layer learning rate, ▽ is the mathematical symbol for gradient, are the LSTM layer parameters, is the loss function, is the updated fully connected layer parameter, is the learning rate of the fully connected layer, is the fully connected layer parameter, is the parameter value after initial training, is the parameter value of the LSTM layer after initial training, is the parameter value of the fully connected layer after initial training.

[0026] Furthermore, the sliding average correction factor is introduced into the dynamic threshold setting, specifically:

[0027]

[0028] in, is the current dynamic threshold, is the smoothing coefficient, is the threshold value at the previous moment, is the sliding window length, i is the intermediate variable, is the historical residual value, is the window residual mean.

[0029] Furthermore, the calculation formula of the historical residual value is as follows:

[0030]

[0031] in, is the historical residual value, for The true value of the moment, for The predicted value at time,i is an intermediate variable;

[0032] The calculation formula of the window residual mean is:

[0033]

[0034] in, is the window residual mean, is the sliding window length.

[0035] In a second aspect, the present invention provides a GNSS real-time deception interference detection system based on deep learning, comprising:

[0036] The training data acquisition module is used to collect the Doppler frequency shift signal of the GNSS receiver in real time to obtain the initial training data set;

[0037] The prediction model training module is used to build an LSTM prediction model with an incremental learning mechanism. Using the initial training dataset, Gaussian white noise is added to the LSTM prediction model with the incremental learning mechanism for model training. The trained LSTM prediction model with the incremental learning mechanism is obtained. The LSTM prediction model with the incremental learning mechanism includes a two-layer LSTM network and a fully connected layer. The two-layer LSTM network introduces a gating mechanism.

[0038] The prediction module is used to extract multidimensional feature vectors after sliding window processing of the initial training data set; the multidimensional feature vectors are input into the trained LSTM prediction model with incremental learning mechanism for prediction to obtain the model prediction value;

[0039] A dynamic threshold setting module is used to calculate the residual between the model prediction value and the real value of the Doppler frequency shift signal collected in real time, calculate the standard deviation of the residual after excluding the warm-up period, and dynamically set the threshold based on the standard deviation of the residual;

[0040] The signal interference event determination module is used to calculate the absolute value of the residual. When the absolute value of the residual exceeds a threshold and the sampling point processed by the sliding window is within a preset interference time window, it is determined that a signal interference event has occurred.

[0041] In a third aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements a deep learning-based GNSS real-time deception interference detection method when executing the computer program.

[0042] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program, and a deep learning-based GNSS real-time deception interference detection method when the computer program is executed by a processor.

[0043] Compared with the prior art, the present invention has the following beneficial technical effects:

[0044] This paper proposes a deep learning-based real-time GNSS spoofing jamming detection method that incorporates an incremental learning mechanism, enabling the model to dynamically adjust parameters based on new data and continuously optimize performance. A two-layer LSTM network, coupled with a gating mechanism, captures complex signal characteristics and long-term dependencies, improving prediction accuracy. Gaussian white noise is added during training to enhance the model's robustness to noise interference, while a hierarchical update strategy with weight freezing balances the impact of historical and new data, preventing drastic parameter fluctuations. Sliding window processing extracts multidimensional feature vectors to comprehensively describe the signal and improve detection sensitivity. Interference events are determined by combining the absolute value of the residual with the interference time window, mitigating random interference. Sliding window processing adapts to spatiotemporal signal variations, while real-time data fusion enables the model to quickly adapt to new environments, achieving efficient prediction and computation, meeting real-time detection requirements, and ensuring stable GNSS system operation. This paper combines LSTM (Long Short-Term Memory) networks with GNSS signal jamming detection to monitor spoofing jamming events in Doppler frequency shift signals from GNSS receivers in real time. The innovative introduction of incremental learning mechanism into LSTM model training solves the performance degradation problem of traditional methods in long time series prediction. Through multi-dimensional feature fusion and hierarchical learning rate setting, the technical effect of interference detection response time less than 0.1 second and detection accuracy greater than 98% is achieved, which significantly improves the GNSS anti-interference capability and realizes real-time deception interference detection of GNSS signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings described herein are for illustrative purposes only and are not intended to limit the scope of the present invention in any way. In addition, the shapes and proportional dimensions of the components in the drawings are only schematic and are used to help understand the present invention, and are not intended to specifically limit the shapes and proportional dimensions of the components of the present invention. In the drawings:

[0046] Figure 1 This is a flowchart of the GNSS real-time deception interference detection method based on deep learning of the present invention.

[0047] Figure 2 This is a structural diagram of the GNSS real-time deception interference detection system based on deep learning of the present invention.

[0048] Figure 3 This is a diagram of the electronic device of the deep learning-based GNSS real-time deception interference detection method of the present invention.

[0049] Figure 4 The figure shows the overall detection flow chart.

[0050] Figure 5 Comparison diagram of input features and original signals.

[0051] Figure 6 This is a comparison chart of traditional LSTM model-based signal detection and residuals.

[0052] Figure 7 This is a signal detection and residual comparison diagram of the LSTM prediction model with incremental learning mechanism of the present invention.

[0053] Figure 8 This is the real-time detection window diagram. DETAILED DESCRIPTION

[0054] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0055] Example 1

[0056] See also Figure 1 , a GNSS real-time deception interference detection method based on deep learning, including the following steps:

[0057] Collect the Doppler frequency shift signal of the GNSS receiver in real time to obtain the initial training data set;

[0058] An LSTM prediction model with an incremental learning mechanism was constructed. Using the initial training dataset, Gaussian white noise was added to the LSTM prediction model with the incremental learning mechanism for model training. The trained LSTM prediction model with the incremental learning mechanism was obtained. The LSTM prediction model with the incremental learning mechanism included a two-layer LSTM network and a fully connected layer. A gating mechanism was introduced into the two-layer LSTM network.

[0059] After performing sliding window processing on the initial training data set, a multidimensional feature vector is extracted; the multidimensional feature vector is input into the trained LSTM prediction model with an incremental learning mechanism for prediction to obtain the model prediction value;

[0060] Calculate the residual between the model prediction value and the real value of the Doppler frequency shift signal collected in real time, calculate the standard deviation of the residual after excluding the warm-up period, and dynamically set the threshold based on the standard deviation of the residual;

[0061] The absolute value of the residual is calculated. When the absolute value of the residual exceeds a threshold and the sampling point processed by the sliding window is within a preset interference time window, it is determined that a signal interference event has occurred.

[0062] This embodiment employs an LSTM prediction model with an incremental learning mechanism. This model can dynamically adjust model parameters and continuously optimize model performance as new data arrives, thereby more accurately predicting the changing trends of Doppler shift signals and improving the ability to model normal signal patterns. This avoids the problem of traditional LSTM models, where parameters are fixed after training and they are difficult to adapt to the dynamic changes in the GNSS signal environment. The two-layer LSTM network increases the model's depth, enabling it to learn more complex signal features. Furthermore, the introduced gating mechanism effectively controls the flow of information, preventing vanishing or exploding gradients, further improving the model's ability to capture long-term dependencies and enhancing prediction accuracy. A hierarchical update strategy with weight freezing is used to dynamically update the prediction model parameters. This not only preserves the model's learning from historical data but also allows it to promptly incorporate features from new data, avoiding drastic fluctuations in model parameters and enhancing the model's stability under varying data distributions. After sliding window processing of the initial training dataset, multidimensional feature vectors are extracted. These feature vectors contain various information about the Doppler shift signal at different time scales, such as mean, variance, and spectral characteristics. The use of multidimensional features can more comprehensively describe signal characteristics, enabling the model to capture more subtle signal changes, thereby improving the detection sensitivity of spoofing interference signals. The method of this embodiment not only considers whether the absolute value of the residual exceeds the threshold, but also requires that the sampling point be within a preset interference time window before determining that a signal interference event has occurred. This comprehensive judgment mechanism avoids false detections due to occasional noise or signal fluctuations, while also reducing missed detections due to the short duration of the interference signal, thereby improving detection reliability. Sliding window processing is applied to the initial training dataset to adapt to the variations of GNSS signals under different temporal and spatial conditions. The sliding window can capture local characteristics of the signal in different time periods, enabling the model to better adapt to the complex and changing GNSS signal environment. The LSTM prediction model can quickly predict the input multidimensional feature vector, calculate the residual sequence between the model prediction and the true value, and dynamically set the threshold. The entire detection process can be completed in a short time, meeting the requirements of real-time GNSS signal spoofing interference detection, enabling timely detection and response to interference events, and ensuring the normal operation of the GNSS system.

[0063] See also Figure 4 , the overall detection flow chart of the deep learning-based GNSS real-time spoofing jamming detection method. The specific steps include:

[0064] Step 1. Collect the Doppler frequency shift signal of the GNSS receiver in real time with a sampling rate of 1000 Hz.

[0065] Step 2. Perform sliding window processing on the signal, with the window length set to 15-30 sampling points, and extract multidimensional feature vectors including instantaneous value, sliding mean, standard deviation, linear trend coefficient, and fast Fourier transform frequency domain component. Step 2 specifically includes:

[0066] The linear trend coefficient is extracted using the least squares fitting method, which is specifically implemented as follows:

[0067]

[0068] Where, trend is the linear trend coefficient, is the sliding window length, For the window signal sampling values, is the time index within the window, i is an intermediate variable.

[0069] Step 3. Build a prediction model consisting of a two-layer LSTM network and a fully connected layer, where the number of nodes in the LSTM hidden layer is 128-512. Step 3 specifically includes:

[0070] The two-layer LSTM network constructed in the model is specifically implemented as follows:

[0071] The input gate is represented as:

[0072] Where, is the input gate activation vector, is the sigmoid activation function, is the weight matrix of the input state, is the weight matrix of the hidden state, is the weight matrix of the cell state, is the characteristic vector of the current input (including instantaneous value, mean, standard deviation, etc.), is the cell state at the previous moment.

[0073] The forget gate is expressed as:

[0074] Where, For the Gate of Forgetfulness, is the hidden state of the model at the previous moment, , are the weight matrices of the current output and current state of the forget gate respectively, is the bias term of the forget gate.

[0075] Cell status update:

[0076] Where, represents the candidate cell state, is the weight matrix currently input to the candidate cell state, is the weight matrix from hidden state to candidate cell state, is the bias term of the candidate cell state, is the hyperbolic tangent activation function.

[0077] final:

[0078] Where, is the cell state, For the Gate of Forgetfulness, represents the Hadamard product, is the cell state at the previous moment, is the input gate activation vector, Represents candidate cell states.

[0079] The output gate is represented as:

[0080] Where, is the output gate, is the sigmoid activation function, is the weight matrix of the current input to the output gate, is the feature vector of the current input, is the weight matrix from the hidden state to the output gate, is the hidden state of the model at the previous moment, is the weight matrix from cell state to output gate, is the cell state, is the bias term of the output gate.

[0081] The hidden state is represented as:

[0082] Where, Indicates the hidden state, is the output gate, represents the Hadamard product, is the hyperbolic tangent activation function, The cell state.

[0083] The two-layer LSTM network in step 3 is constructed using the above formula, with an input dimension of 5 (corresponding to 5 features) and a hidden layer dimension of 256.

[0084] Step 4. Use the initial training dataset to train the model. The training rounds are 8-15 epochs, and the initial learning rate is set to 0.0005-0.005.

[0085] Step 5. When new monitoring data is received, 5%-15% of historical training data is randomly selected and mixed with the new data. A hierarchical learning rate strategy is used to update model parameters, with the LSTM layer learning rate being 0.1-0.3 times that of the fully connected layer, and training lasting 3-8 epochs. New monitoring data refers to data that differs from the data used during model training. Step 5 employs a hierarchical update strategy with weight freezing. Specifically, weight updates for the LSTM layer are limited to ±10% of the initial weights; weight updates for the fully connected layer are limited to ±20% of the initial weights; and gradient clipping is employed, with a gradient norm threshold set to 1.0-2.0. The hierarchical learning rate is an optimization strategy, and weight freezing is a parameter constraint mechanism. Together, they address model degradation in dynamic GNSS signal environments.

[0086] The incremental learning rule for layered updates is:

[0087]

[0088] The constraints are:

[0089]

[0090]

[0091] Where, is the updated LSTM layer parameter, is the LSTM layer learning rate, ▽ is the mathematical symbol for gradient, are the LSTM layer parameters, is the loss function, is the updated fully connected layer parameter, is the learning rate of the fully connected layer, is the fully connected layer parameter, is the parameter value after initial training, is the parameter value of the LSTM layer after initial training, is the parameter value of the fully connected layer after initial training.

[0092] Step 6. Calculate the residual sequence between the model prediction value and the true value, dynamically calculate the residual standard deviation after excluding the warm-up period of the first 50-100 sampling points, and set the threshold to 5-10 times the standard deviation. The true value is the new monitoring data received. Each time, the model predicts the next data based on the data in the window, and then compares (differences) it with the received true data to determine whether it is interfered with. The dynamic threshold calculation in step 6 introduces a sliding average correction factor, which can automatically adjust according to the actual changes in the signal to reduce the false detection and missed detection rates. The specific calculation formula is:

[0093]

[0094] Where, is the current dynamic threshold, is the threshold value at the previous moment, is the smoothing coefficient, is the sliding window length, i is the intermediate variable, is the historical residual value, is the mean of the window residuals;

[0095] The formula for calculating historical residual value is:

[0096] Where, for The true value of the moment, for The predicted value at time.

[0097] The formula for calculating the window residual mean is: .

[0098] Step 7: When the absolute value of the residual of the sampling point exceeds the threshold and is within the preset interference time window, it is determined that a signal interference event occurs.

[0099] The LSTM model with incremental learning mechanism proposed in this embodiment aims to solve the problem of inaccurate prediction of traditional LSTM models in long time series.

[0100] The specific implementation flow chart is as follows Figure 4 shown.

[0101] In the data acquisition and preprocessing stage, the original data is the intermediate frequency Doppler shift signal output by the GNSS receiver, the sampling rate is set to fs = 1000 Hz, and the data format is a one-dimensional time series { }, where t=1,2…T. In addition, Gaussian white noise is added during the training phase to enhance the robustness of the model. The Gaussian white noise formula is:

[0102]

[0103] Where: is noise, is the noise intensity coefficient, is the standard deviation of the original signal, is a standard normal distribution.

[0104] After the model is started, it first checks whether there is an initially trained model. If so, it loads the saved model and immediately starts the real-time data stream module to process the new data in real time. If there is no saved model, it performs initial training and saves the model and parameters after the training is completed.

[0105] See also Figure 5 , comparison chart of input features and original signals, Figure 5 a is the original signal comparison, Figure 5 b is the current value feature comparison, Figure 5 c is the sliding mean feature comparison, Figure 5 d is the standard deviation feature contrast, Figure 5 The e is the trend slope characteristic comparison, Figure 5 f represents the FFT (Fast Fourier Transform) dominant frequency amplitude feature comparison. When training the model using the initial training dataset, the training rounds are 8-15 epochs, and the initial learning rate is set to 0.0005-0.005. First, multi-dimensional feature extraction is performed on the initial training data, with a sliding window length set to 15-30 sampling points. Multi-dimensional feature vectors are extracted, including instantaneous values, sliding means, standard deviations, linear trend coefficients, and fast Fourier transform frequency domain components. The feature extraction formula is as follows:

[0106] The instantaneous value is expressed as:

[0107] The mean is expressed as:

[0108] The standard deviation is expressed as:

[0109] The trend term is expressed as:

[0110] The frequency domain components are expressed as:

[0111] In the above formula is the time index within the window, is the number of data in the window, The value of the nth data in the window.

[0112] The multi-dimensional feature vector of the real-time data stream is extracted from the data buffer in the above way, and the extracted feature vector is input into the LSTM prediction model with incremental learning mechanism for prediction. The LSTM prediction model with incremental learning mechanism mainly consists of an input layer (containing a 5-dimensional feature vector), an LSTM layer (the number of hidden nodes is 256), and an output layer (outputting a one-dimensional prediction value). ).

[0113] When new data is received, the following steps are performed to dynamically update the model:

[0114] Step 1: Data mixing: Take all samples from the new dataset ; Randomly extract 10% of the samples from the historical data ; Merge datasets: .

[0115] Step 2: Layer parameter update: Set the learning rate of the LSTM layer to 0.1 times that of the fully connected layer;

[0116] The optimization goal is to minimize the mean square error:

[0117]

[0118] The parameter update formula is:

[0119]

[0120]

[0121] Where, for The true value of the moment, for The predicted value at time. are the LSTM layer parameters, is the fully connected layer parameter, is the LSTM layer learning rate, is the learning rate of the fully connected layer, is the loss function, , and uses gradient clipping technology to limit the amplitude of parameter changes (gradient norm threshold 1.0).

[0122] Dynamic threshold interference judgment: After excluding the first 50-100 sampling points during the warm-up period, the first 50-100 sampling points during the warm-up period are excluded and do not participate in the subsequent residual calculation or threshold setting. The residual standard deviation is calculated and the threshold is set dynamically:

[0123]

[0124]

[0125] Where: is the dynamic threshold, is the residual standard deviation, is the adjustment coefficient, is the number of samples in the warm-up period, is the preset interference detection time window length, is the prediction residual, is the window residual mean, and the calculation formula is:

[0126]

[0127] Where m is the sliding window length.

[0128] Interference judgment logic: The interference alarm is triggered when the following conditions are met at the same time:

[0129] (1) The absolute value of the residual of the sampling point exceeds the threshold;

[0130] (2) The abnormal point is located within the preset interference time window;

[0131] (3) Statistical difference D ≥ 2.5.

[0132] See also Figure 6 , traditional LSTM model-based signal detection and residual comparison chart. Comparison of the traditional LSTM model-based GNSS signal prediction results and the real signal. The results show that when the prediction time step exceeds the length range of the training signal, the predicted signal trend tends to be flat, and deviates significantly from the dynamic changes of the real signal. As the time step increases, the residual gradually increases, especially at the end of the long sequence, the residual peak reaches more than 5 times that of the training stage. The reason is that the traditional LSTM model lacks a dynamic update mechanism and has limited long-term dependence capabilities, and cannot adapt to the non-stationary characteristics of GNSS signals. This embodiment verifies the limitations of static models in real-time detection and provides a basis for introducing incremental learning strategies.

[0133] See also Figure 7 , the signal detection and residual comparison diagram of the improved LSTM prediction model with incremental learning mechanism. This shows the detection effect of the improved LSTM prediction model with incremental learning mechanism in this embodiment. The key observations are as follows:

[0134] The predicted curve closely matches the time-frequency characteristics of the real signal, especially at sudden changes in the signal, where the model quickly tracks the changes. The residual error remains low, and the residual spikes at interference points are accurately captured. The model continuously adapts to dynamic signal changes, resolving the long-sequence prediction degradation problem inherent in traditional LSTM models.

[0135] See also Figure 8 The real-time detection window diagram shows the real-time judgment process within the dynamic detection window. The model-predicted signal deviates significantly from the true value within the interference window. During the interference period, the residual error continuously exceeds the dynamic threshold, triggering a deception jamming alarm.

[0136] The effect verification of the embodiment is shown in Table 1:

[0137] Table 1 Final operation result evaluation table

[0138]

[0139] The F1 score in Table 1 is an indicator used in statistics to measure the accuracy of the model.

[0140] After running the program, the model's performance was tested by varying the sliding window length from 20 to 200, using scenarios from the Texas Spoofing Test Battery (TEXBAT) dataset. The results showed detection accuracy exceeding 98%, with a response time within 0.1 seconds, sufficient for subsequent engineering applications. The TEXBAT dataset, a multi-scenario spoofing jamming dataset released by the University of Texas, is one of the most widely used and systematically structured GNSS spoofing test sets.

[0141] This embodiment implements interference detection using an LSTM prediction model with an incremental learning mechanism. The LSTM model excels at processing and predicting time series data. Its built-in memory mechanism can capture dynamic changes and long-term dependencies of signals over time. This makes the LSTM model outperform DNN (Deep Neural Network) models in tasks involving time series correlation or signal variation patterns. Furthermore, the LSTM network incorporates a gating mechanism, including an input gate, a forget gate, and an output gate, which selectively memorizes or forgets past information. This is particularly helpful for identifying important features in time series patterns and ignoring noise. This embodiment calculates the residual sequence between the model's predicted value and the true value of the Doppler frequency shift signal acquired in real time. After excluding the warm-up period, the standard deviation of the residual is calculated and a threshold is dynamically set. A sliding average correction factor is also incorporated into the threshold calculation. This dynamic threshold automatically adjusts based on actual signal changes, avoiding the potential for false or missed detections caused by fixed thresholds when signals fluctuate significantly, thereby improving detection accuracy. Gaussian white noise is added to the model during training to simulate various interference factors that may exist in real-world environments. This training method makes the model more stable in the face of noise interference, better able to extract effective signal features from noise, and improves the model's adaptability and robustness to actual GNSS signals. Randomly selecting a portion of historical training data and mixing it with the real-time Doppler shift signal for prediction enables the model to promptly incorporate new real-time data, quickly adapting to changes in the signal environment, and improving the model's adaptability in complex environments.

[0142] The LSTM model can be well applied to different types of time series data. It has strong adaptability and can handle dynamic changes and nonlinear relationships in signals. It can effectively capture and understand complex time series data patterns, long-term dependencies, and the time-varying characteristics of interference signals. In particular, in this embodiment, incremental learning is combined with the LSTM model, so that the model can be updated online and perform real-time signal processing.

[0143] Example 2

[0144] See also Figure 2, a deep learning-based GNSS real-time spoofing and jamming detection system, including:

[0145] The training data acquisition module is used to collect the Doppler frequency shift signal of the GNSS receiver in real time to obtain the initial training data set;

[0146] The prediction model training module is used to build an LSTM prediction model with an incremental learning mechanism. Using the initial training dataset, Gaussian white noise is added to the LSTM prediction model with the incremental learning mechanism for model training. The trained LSTM prediction model with the incremental learning mechanism is obtained. The LSTM prediction model with the incremental learning mechanism includes a two-layer LSTM network and a fully connected layer. The two-layer LSTM network introduces a gating mechanism.

[0147] The prediction module is used to extract multidimensional feature vectors after sliding window processing of the initial training data set; the multidimensional feature vectors are input into the trained LSTM prediction model with incremental learning mechanism to obtain the model prediction value;

[0148] A dynamic threshold setting module is used to calculate the residual between the model prediction value and the real value of the Doppler frequency shift signal collected in real time, calculate the standard deviation of the residual after excluding the warm-up period, and dynamically set the threshold based on the standard deviation of the residual;

[0149] The signal interference event determination module is used to calculate the absolute value of the residual. When the absolute value of the residual exceeds a threshold and the sampling point processed by the sliding window is within a preset interference time window, it is determined that a signal interference event has occurred.

[0150] Example 3

[0151] See also Figure 3 , an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements a deep learning-based GNSS real-time deception interference detection method when executing the computer program.

[0152] Example 4

[0153] A computer-readable storage medium stores a computer program, and the computer program is executed by a processor to implement a deep learning-based GNSS real-time spoofing interference detection method.

[0154] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, read-only optical disks, optical storage devices, etc.) containing computer-usable program code.

[0155] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0156] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0157] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the specific implementations of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be included in the scope of protection of the present invention.

Claims

1. A deep learning-based GNSS real-time deception jamming detection method, characterized by: The following steps are involved: Collect the Doppler frequency shift signal of the GNSS receiver in real time to obtain the initial training data set; An LSTM prediction model with an incremental learning mechanism was constructed. Using the initial training dataset, Gaussian white noise was added to the LSTM prediction model with the incremental learning mechanism for model training. The trained LSTM prediction model with the incremental learning mechanism was obtained. The LSTM prediction model with the incremental learning mechanism included a two-layer LSTM network and a fully connected layer. A gating mechanism was introduced into the two-layer LSTM network. After performing sliding window processing on the initial training data set, a multidimensional feature vector is extracted. The multidimensional feature vector is then input into a trained LSTM prediction model with an incremental learning mechanism to obtain a model prediction value. The prediction is specifically performed by randomly selecting a portion of historical training data and mixing it with the real-time collected Doppler frequency shift signal, and dynamically updating the prediction model parameters using a hierarchical update strategy with weight freezing. The dynamic update of the prediction model parameters using the hierarchical update strategy with weight freezing includes a gradient norm threshold and a hierarchical weight update amount constraint, and the gradient norm threshold and the hierarchical weight update amount constraint include: The gradient norm threshold range is set using the gradient clipping method; The weight update amount of the two-layer LSTM network is within the first threshold range of the initial weight; The weight update amount of the fully connected layer is within the second threshold range of the initial weight; The hierarchical update strategy with weight freezing adopts an incremental learning rule, which is specifically: The constraints of the incremental learning rule are: in, is the updated LSTM layer parameter, is the LSTM layer learning rate, ▽ is the mathematical symbol for gradient, are the LSTM layer parameters, is the loss function, is the updated fully connected layer parameter, is the learning rate of the fully connected layer, is the fully connected layer parameter, is the parameter value after initial training, is the parameter value of the LSTM layer after initial training, is the parameter value of the fully connected layer after initial training; Calculate the residual between the model prediction value and the real value of the Doppler frequency shift signal collected in real time, calculate the standard deviation of the residual after excluding the warm-up period, and dynamically set the threshold based on the standard deviation of the residual; The absolute value of the residual is calculated. When the absolute value of the residual exceeds a threshold and the sampling point processed by the sliding window is within a preset interference time window, it is determined that a signal interference event has occurred.

2. The GNSS real-time deception interference detection method based on deep learning according to claim 1 is characterized in that: The dynamic threshold setting introduces a sliding average correction factor, specifically: in, is the current dynamic threshold, is the smoothing coefficient, is the threshold value at the previous moment, is the sliding window length, i is an intermediate variable, is the historical residual value, is the window residual mean.

3. The deep learning-based GNSS real-time deception interference detection method according to claim 2, characterized in that: The calculation formula of the historical residual value is specifically: in, is the historical residual value, for The true value of the moment, for The predicted value at time, i is an intermediate variable; The calculation formula of the window residual mean is specifically: in, is the window residual mean, is the sliding window length.

4. A deep learning-based GNSS real-time deception jamming detection system, characterized by: The method for detecting GNSS real-time deception interference based on deep learning as described in any one of claims 1 to 3 comprises: The training data acquisition module is used to collect the Doppler frequency shift signal of the GNSS receiver in real time to obtain the initial training data set; The prediction model training module is used to build an LSTM prediction model with an incremental learning mechanism. Using the initial training dataset, Gaussian white noise is added to the LSTM prediction model with the incremental learning mechanism for model training. The trained LSTM prediction model with the incremental learning mechanism is obtained. The LSTM prediction model with the incremental learning mechanism includes a two-layer LSTM network and a fully connected layer. The two-layer LSTM network introduces a gating mechanism. The prediction module is used to extract multidimensional feature vectors after sliding window processing of the initial training data set; the multidimensional feature vectors are input into the trained LSTM prediction model with incremental learning mechanism for prediction to obtain the model prediction value; A dynamic threshold setting module is used to calculate the residual between the model prediction value and the real value of the Doppler frequency shift signal collected in real time, calculate the standard deviation of the residual after excluding the warm-up period, and dynamically set the threshold based on the standard deviation of the residual; The signal interference event determination module is used to calculate the absolute value of the residual. When the absolute value of the residual exceeds a threshold and the sampling point processed by the sliding window is within a preset interference time window, it is determined that a signal interference event has occurred.

5. An electronic device, characterized in that: The invention comprises a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for detecting GNSS real-time deception interference based on deep learning as described in any one of claims 1 to 3 is implemented.

6. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the deep learning-based GNSS real-time deception interference detection method described in any one of claims 1 to 3.

Images