Small power GNSS interference recognition method based on convolutional neural network
By constructing a hierarchical recognition classifier based on convolutional neural networks and combining time-frequency images, autocorrelation functions, and spectral flatness features, the problem of low accuracy in low-power GNSS interference recognition was solved, and efficient interference recognition was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CIVIL AVIATION UNIV OF CHINA
- Filing Date
- 2024-05-30
- Publication Date
- 2026-06-16
AI Technical Summary
Existing GNSS interference identification methods have low accuracy under low-power interference conditions, and existing models are not ideal in identifying band-limited Gaussian noise interference and narrowband noise FM interference, and are highly complex.
A hierarchical recognition classifier based on convolutional neural networks is adopted, including a time-frequency image classifier, an autocorrelation function image classifier, and a spectral flatness classifier. By extracting features from short-time Fourier transform, autocorrelation function, and spectral flatness, and combining channel attention mechanism and residual connection module, the recognition accuracy of low-power interference is improved.
Even at a low interference power of -110dBm, the recognition accuracy reaches 93.29%, with a low false alarm rate, and it does not require a large amount of training data, which significantly improves the recognition accuracy and efficiency of low-power interference.
Smart Images
Figure CN118484718B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of satellite navigation anti-interference technology, and specifically relates to a low-power Global Navigation Satellite System (GNSS) interference identification method based on convolutional neural networks, which can achieve an interference identification accuracy of 93.29% even when the interference power is -110dBm. Background Technology
[0002] GNSS provides precise positioning, navigation, and timing services, thus playing a crucial role in transportation, communication, and military fields. However, navigation satellites are far from the ground, and GNSS signals are very weak when they reach the ground, making them highly susceptible to interference. Previous research to improve GNSS navigation performance in interference environments has primarily focused on developing effective anti-interference methods for specific types of interference. However, the actual operating environment of GNSS receivers is complex and variable, and interference identification can provide a basis for targeted selection of anti-interference measures.
[0003] Existing interference identification methods perform well in identifying high-power interference (greater than -105 dBm), but their accuracy drops significantly for low-power interference (less than -105 dBm). Given that the power of satellite signals reaching GNSS receivers is typically around -125 dBm, this means that the impact of low-power interference on satellite signals is not negligible. Furthermore, existing research employs complex models to distinguish between band-limited Gaussian noise interference and interference-free conditions, and the identification accuracy for spoofing interference and narrowband noise FM interference remains unsatisfactory. Summary of the Invention
[0004] To address the aforementioned problems, the present invention aims to provide a GNSS interference identification method based on a convolutional neural network, which can achieve a high interference identification accuracy even when the interference power is low.
[0005] To achieve the above objectives, the low-power GNSS interference identification method based on convolutional neural networks provided by this invention includes the following steps performed in sequence:
[0006] 1) Generate 9 types of GNSS interference signals to form an interference signal dataset, and then divide the interference signal dataset into training set and test set according to the proportion;
[0007] 2) Construct a hierarchical recognition classifier consisting of a time-frequency image classifier, an autocorrelation function image classifier, and a spectral flatness classifier;
[0008] 3) Use the training set obtained in step 1) to train the hierarchical recognition classifier constructed in step 2) to obtain the trained hierarchical recognition classifier;
[0009] 4) Input the test set obtained in step 1) into the trained hierarchical recognition classifier obtained in step 3) for testing and verification to obtain the final hierarchical recognition classifier;
[0010] 5) Input the real-time acquired GNSS interference signal into the final hierarchical identification classifier obtained in step 4), and the final hierarchical identification classifier outputs the category of the GNSS interference signal.
[0011] In step 1), the nine types of GNSS interference signals are single-tone interference, multi-tone interference, linear chirp interference, sinusoidal chirp interference, pulse interference, narrowband noise FM interference, spoofing interference, band-limited Gaussian noise interference, and no interference. The interference signal dataset is composed of all GNSS interference signals, and then the interference signal dataset is divided into a training set and a test set at a ratio of 10:1.
[0012] The power of each GNSS jamming signal is in the range of [-110, -101] dBm, the initial phase of the carrier is in the range of [0, 2π], and other parameters of the jamming signal are set as follows: the carrier frequency of single-tone jamming is in the range of [0, 2] MHz; multi-tone jamming includes 2 single-tone jammings, each with a carrier frequency in the range of [0, 2] MHz; the sweep bandwidth of linear chirp jamming and sinusoidal chirp jamming is in the range of [0.5, 2] MHz, and the sweep period is in the range of [20, 50] μs; the bandwidth of narrowband noise FM jamming is in the range of [0.1, 2] MHz; the bandwidth of band-limited Gaussian noise jamming is in the range of [0.5, 2] MHz; the repetition period of pulse jamming is in the range of [0.02, 0.2] ms; and the Doppler frequency shift of spoofing jamming is in the range of [0, 2] MHz.
[0013] In step 2), the time-frequency image classifier includes: a first convolutional layer, a second convolutional layer, a first max-pooling layer, a third convolutional layer, a second max-pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a Softmax layer; a channel attention mechanism is used after the second max-pooling layer; a residual connection module is used between the first and second convolutional layers; a residual connection module is used between the first and third convolutional layers; the kernel filter size of all convolutional layers and max-pooling layers is set to 3×3; the number of kernel filters in the first and second convolutional layers is 32; the number of kernel filters in the third convolutional layer is 64; the stride of the kernel filters in the convolutional layers is 1; the stride of the kernel filters in the max-pooling layers is 2; the activation function of all convolutional layers is set to the ReLU activation function; and the number of neurons in the first, second, and third fully connected layers is set to 512, 256, and 7, respectively.
[0014] The autocorrelation function image classifier includes: a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a softmax layer; the kernel filter size of all convolutional layers and max pooling layers is set to 3×3, the number of kernel filters in the first convolutional layer is 32, the number of kernel filters in the second convolutional layer is 64, the stride of the kernel filters in the convolutional layers is 1, the stride of the kernel filters in the max pooling layers is 2, and the number of neurons in the first, second, and third fully connected layers is set to 512, 256, and 2, respectively.
[0015] In step 3), the training set obtained in step 1) is used to train the hierarchical recognition classifier constructed in step 2), and the specific method for obtaining the trained hierarchical recognition classifier is as follows:
[0016] Input the training set obtained in step 1) into the hierarchical recognition classifier constructed in step 2), train the hierarchical recognition classifier, set the initial learning rate to 0.0001, select the Adam algorithm as the optimization algorithm, and select the cross-entropy loss function as the loss function.
[0017] First, a time-frequency image classifier performs a short-time Fourier transform on the GNSS interference signals in the training set, and then denoises the time-frequency transform results to obtain a denoised time-frequency image of the GNSS interference signal as the first layer of features. If the interference is single-tone, multi-tone, linear chirp, sinusoidal chirp, or impulse, the time-frequency image classifier directly outputs the recognition result. If the interference is spoofing or narrowband noise FM interference, the autocorrelation function image classifier calculates the autocorrelation function of the GNSS interference signal and, based on the autocorrelation function... The autocorrelation function image is plotted using the autocorrelation function as the second layer feature. The autocorrelation function image is then used to perform secondary identification of GNSS interference signals, thereby distinguishing between deception interference and narrowband noise FM interference. If it is band-limited Gaussian noise interference or no interference, a fast Fourier transform is performed on the GNSS interference signal using a spectral flatness classifier to obtain the spectral flatness of the GNSS interference signal as the third layer feature. The spectral flatness is then compared with a set threshold. If it is greater than the threshold, it is no interference; if it is less than the threshold, it is band-limited Gaussian noise interference. This process is used to perform secondary identification of GNSS interference signals.
[0018] The trained hierarchical classifier is obtained when the rate of change of the cross-entropy loss function is less than 0.1%.
[0019] The short-time Fourier transform formula is as follows:
[0020]
[0021] Where y(n) is the GNSS interference signal, n is the time variable, and h(nm) is a window function centered at time m, where m = 0, 1, ..., M. y -1, where k is a discrete frequency point, k = 0, 1, ..., N y -1, j is the imaginary unit, STFT(m,k) is the time-frequency transformation result of the GNSS interference signal, N y M is the number of points in the Fast Fourier Transform. y Let w be the number of times the window function slides, where w = 2π / N y .
[0022] The method for denoising the time-frequency transformation results of GNSS interference signals is as follows:
[0023] For the calculated time-frequency transform result STFT(m,k) of the GNSS interference signal, the time-frequency function value corresponding to its third quartile is selected as the threshold λ. m Then, the time-frequency transformation result of the GNSS interference signal is denoised using the following soft threshold denoising method:
[0024]
[0025] Where sgn(STFT(m,k)) is the sign function, expressed as follows:
[0026]
[0027] The autocorrelation function is obtained by the following formula:
[0028]
[0029] Where τ is the time delay and N is the length of the GNSS interference signal y(n).
[0030] The spectral flatness is obtained by the following formula:
[0031]
[0032] Where k is a discrete frequency point, Y(k) is the result of the fast Fourier transform of the GNSS interference signal y(n), and M is the number of discrete frequency points.
[0033] In step 2), the specific operation methods of the channel attention mechanism and residual connection module in the time-frequency image classifier are as follows:
[0034] The channel attention mechanism, through squeezing and activation operations, enables the network to adaptively learn the weights of each channel and adjust the importance of each channel's features in subsequent network layers, thereby improving the classifier's performance. The input to the squeezing operation is a feature map U of size C×H×W. sq =[u1,u2,…,u C ], C is the number of channels in the input feature map, for feature map U sq Perform global average pooling to compress it into a feature vector z = [z1, z2, ..., z] of size C × 1 × 1. C The k-th element z in the eigenvector z k Represented as:
[0035]
[0036] Among them, u k Let H be the feature map of the kth input channel, and let H and W be the length and width of the feature map of that channel, respectively.
[0037] In the activation operation, the feature vector z generated by the squeezing operation is passed through a fully connected layer and a nonlinear activation function to learn and generate the input feature map U. sq The weight vector s = [s1, s2, ..., s C The formula for calculating the weight vector s is as follows:
[0038] s=σ(W2δ(W1z))
[0039] Where W1∈R C'×C W2∈R C×C' C' is less than C; δ is the ReLU activation function, σ is the sigmoid activation function, and finally the learned weight vector s is applied to the input feature map U. sq For each channel, the final output feature map U' is obtained. sq =[u'1,u'2,…,u' C ], where u' k The weighted result of the feature map of the k-th input channel is represented as:
[0040] u' k =s k u k
[0041] The residual connection module skips some network layers from the input features and connects them directly to the output.
[0042] The low-power GNSS interference identification method based on convolutional neural networks provided by this invention has the following advantages:
[0043] 1) This invention performs noise reduction processing on time-frequency images, which improves the quality of time-frequency images and can improve the accuracy of interference recognition under low power conditions.
[0044] 2) In addition to the time-frequency plot, this invention also introduces two features: autocorrelation function image and spectral flatness. The autocorrelation function image can help distinguish between narrowband noise FM interference and spoofing interference, while spectral flatness can help distinguish between band-limited Gaussian noise interference and no interference.
[0045] 3) Different features have varying importance to the classification results. Directly fusing features at the input layer may ignore some important features or give excessive weight to secondary features, leading to a decrease in model performance. This invention adopts a hierarchical recognition strategy. First, a time-frequency image classifier is used to identify interference signals with high time-frequency image discriminativeness. Narrowband noise FM interference and spoofing interference with high time-frequency image similarity are then identified a second time using an autocorrelation function image classifier. Band-limited Gaussian noise interference and interference-free cases are then identified a second time using a spectral flatness classifier. This improves the accuracy of identifying interference signals with low time-frequency image discriminativeness. Attached Figure Description
[0046] Table 1 shows the confusion matrix of the identification method of the present invention when the interference power is -110dBm.
[0047] Figure 1 The flowchart of the low-power GNSS interference identification method based on convolutional neural networks provided by the present invention is shown.
[0048] Figure 2 This is a structural diagram of the hierarchical recognition classifier in this invention.
[0049] Figure 3 This is a structural diagram of the residual connection module in this invention.
[0050] Figure 4(a) is a graph showing the change in recognition accuracy of the method of the present invention under different interference power conditions.
[0051] Figure 4(b) shows the false alarm rate variation curves of the method of the present invention under different interference power conditions.
[0052] Figure 5 This is a graph showing the change in recognition accuracy of the method of the present invention under different training data amounts. Detailed Implementation
[0053] The low-power GNSS interference identification method based on convolutional neural networks provided by the present invention will be described in detail below with reference to the accompanying drawings and embodiments.
[0054] like Figure 1As shown, the low-power GNSS interference identification method based on convolutional neural networks provided by this invention includes the following steps performed in sequence:
[0055] 1) Generate 9 types of GNSS interference signals to form an interference signal dataset, and then divide the interference signal dataset into training set and test set according to the proportion;
[0056] The nine types of GNSS interference signals are single-tone interference, multi-tone interference, linear chirp interference, sinusoidal chirp interference, pulse interference, narrowband noise FM interference, deception interference, band-limited Gaussian noise interference, and no interference. The interference signal dataset is composed of all GNSS interference signals, and then the interference signal dataset is divided into training set and test set at a ratio of 10:1.
[0057] The power of each GNSS jamming signal is in the range of [-110, -101] dBm, the initial phase of the carrier is in the range of [0, 2π], and other parameters of the jamming signal are set as follows: the carrier frequency of single-tone jamming is in the range of [0, 2] MHz; multi-tone jamming includes 2 single-tone jammings, each with a carrier frequency in the range of [0, 2] MHz; the sweep bandwidth of linear chirp jamming and sinusoidal chirp jamming is in the range of [0.5, 2] MHz, and the sweep period is in the range of [20, 50] μs; the bandwidth of narrowband noise FM jamming is in the range of [0.1, 2] MHz; the bandwidth of band-limited Gaussian noise jamming is in the range of [0.5, 2] MHz; the repetition period of pulse jamming is in the range of [0.02, 0.2] ms; and the Doppler frequency shift of spoofing jamming is in the range of [0, 2] MHz.
[0058] In this embodiment, based on the parameters of the nine GNSS interference signals described above, the SIGLENT SSG3032X RF signal generator and SDG6052X-E arbitrary waveform generator are used to generate and transmit different types of GNSS interference signals in the L1 band. These signals are then received and processed using a software receiver based on BladeRF, and L1 band satellite signals generated by a GPS signal simulator are superimposed. The receiver front-end filter bandwidth and sampling frequency are fixed at 2.2MHz and 10MHz, respectively. For each GNSS interference signal, 220 samples are generated in 1dBm steps within each interference power range, resulting in a total of 2200 samples for each GNSS interference signal and a total of 19800 samples in the interference signal dataset.
[0059] 2) such as Figure 2 As shown, a hierarchical recognition classifier is constructed, consisting of a time-frequency image classifier, an autocorrelation function image classifier, and a spectral flatness classifier.
[0060] The time-frequency image classifier includes: a first convolutional layer, a second convolutional layer, a first max-pooling layer, a third convolutional layer, a second max-pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a Softmax layer; a channel attention mechanism is used after the second max-pooling layer; a residual connection module is used between the first and second convolutional layers; a residual connection module is used between the first and third convolutional layers; the kernel filter size of all convolutional layers and max-pooling layers is set to 3×3; the number of kernel filters in the first and second convolutional layers is 32; the number of kernel filters in the third convolutional layer is 64; the stride of the kernel filters in the convolutional layers is 1; the stride of the kernel filters in the max-pooling layers is 2; the activation function of all convolutional layers is set to ReLU; and the number of neurons in the first, second, and third fully connected layers is set to 512, 256, and 7, respectively.
[0061] The specific operation methods of the channel attention mechanism and the residual connection module are as follows:
[0062] The channel attention mechanism, through squeezing and activation operations, enables the network to adaptively learn the weights of each channel and adjust the importance of each channel's features in subsequent network layers, thereby improving the classifier's performance. The input to the squeezing operation is a feature map U of size C×H×W. sq =[u1,u2,…,u C ], C is the number of channels in the input feature map, for feature map U sq Perform global average pooling to compress it into a feature vector z = [z1, z2, ..., z] of size C × 1 × 1. C The k-th element z in the eigenvector z k Represented as:
[0063]
[0064] Among them, u k Let H be the feature map of the kth input channel, and let H and W be the length and width of the feature map of that channel, respectively.
[0065] In the activation operation, the feature vector z generated by the squeezing operation is passed through a fully connected layer and a nonlinear activation function to learn and generate the input feature map U. sq The weight vector s = [s1, s2, ..., s C The formula for calculating the weight vector s is as follows:
[0066] s=σ(W2δ(W1z))
[0067] Where W1∈R C'×C W2∈R C×C'C' is less than C; δ is the ReLU activation function, σ is the sigmoid activation function, and finally the learned weight vector s is applied to the input feature map U. sq For each channel, the final output feature map U' is obtained. sq =[u'1,u'2,…,u' C ], where u' k The weighted result of the feature map of the k-th input channel is represented as:
[0068] u' k =s k u k
[0069] like Figure 3 As shown, the residual connection module skips some network layers and connects the input features directly to the output, solving the gradient vanishing or gradient exploding problem caused by the increase of network depth. It preserves the original input information and ensures that the performance of the next layer is at least the same as that of the current layer, rather than worse. Figure 3 The network, consisting of two superimposed "weight layers," is used to process the input features U. res To residual F(U) res The mapping of ) will result in the residual F(U) res ) and input features U res Perform element-level addition to obtain the output U' of the residual connection module. res .
[0070] The autocorrelation function image classifier includes: a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a softmax layer; the kernel filter size of all convolutional layers and max pooling layers is set to 3×3, the number of kernel filters in the first convolutional layer is 32, the number of kernel filters in the second convolutional layer is 64, the stride of the kernel filters in the convolutional layers is 1, the stride of the kernel filters in the max pooling layers is 2, and the number of neurons in the first, second, and third fully connected layers is set to 512, 256, and 2, respectively.
[0071] 3) Use the training set obtained in step 1) to train the hierarchical recognition classifier constructed in step 2) to obtain the trained hierarchical recognition classifier;
[0072] The specific method is as follows:
[0073] Input the training set obtained in step 1) into the hierarchical recognition classifier constructed in step 2), train the hierarchical recognition classifier, set the initial learning rate to 0.0001, select the Adam algorithm as the optimization algorithm, and select the cross-entropy loss function as the loss function.
[0074] First, the GNSS interference signals in the training set are transformed using the Short Time Fourier Transform (STFT) through a time-frequency image classifier, and the time-frequency transformation result is denoised to obtain the denoised time-frequency image of the GNSS interference signal as the first layer of features.
[0075] The short-time Fourier transform formula is as follows:
[0076]
[0077] Where y(n) is the GNSS interference signal, n is the time variable, and h(nm) is a window function centered at time m, where m = 0, 1, ..., M. y -1, where k is a discrete frequency point, k = 0, 1, ..., N y -1, j is the imaginary unit, STFT(m,k) is the time-frequency transformation result of the GNSS interference signal, N y M is the number of points in the Fast Fourier Transform. y Let w be the number of times the window function slides, where w = 2π / N y .
[0078] The method for denoising the time-frequency transformation results is as follows:
[0079] For the calculated time-frequency transform result STFT(m,k) of the GNSS interference signal, the time-frequency function value corresponding to its third quartile is selected as the threshold λ. m Then, the time-frequency transformation result of the GNSS interference signal is denoised using the following soft threshold denoising method:
[0080]
[0081] Where sgn(STFT(m,k)) is the sign function, expressed as follows:
[0082]
[0083] Time-frequency images can effectively identify five types of interference signals: single-tone interference, multi-tone interference, linear chirp interference, sinusoidal chirp interference, and impulse interference. However, they are prone to confusing narrowband noise FM interference with spoofing interference, and band-limited Gaussian noise interference with no interference. Therefore, for single-tone interference, multi-tone interference, linear chirp interference, sinusoidal chirp interference, and impulse interference, the time-frequency image classifier directly outputs the identification result. For spoofing interference and narrowband noise FM interference, the autocorrelation function image classifier calculates the autocorrelation function of the GNSS interference signal and plots the autocorrelation function image as the second layer of features. The autocorrelation function image is then used for secondary identification of the GNSS interference signal, thereby distinguishing between spoofing interference and narrowband noise FM interference.
[0084] The autocorrelation function is obtained by the following formula:
[0085]
[0086] Where τ is the time delay and N is the length of the GNSS interference signal y(n).
[0087] If the interference is band-limited Gaussian noise or no interference, the GNSS interference signal is subjected to Fast Fourier Transform (FFT) by a spectral flatness classifier to obtain the spectral flatness of the GNSS interference signal as the third layer feature. The spectral flatness is then compared with a set threshold. If it is greater than the threshold, it is an interference-free case; if it is less than the threshold, it is a band-limited Gaussian noise interference. This process is used to identify the GNSS interference signal a second time.
[0088] The spectral flatness is obtained by the following formula:
[0089]
[0090] Where k is a discrete frequency point, Y(k) is the result of the fast Fourier transform of the GNSS interference signal y(n), and M is the number of discrete frequency points.
[0091] The trained hierarchical classifier is obtained when the rate of change of the cross-entropy loss function is less than 0.1%.
[0092] 4) Input the test set obtained in step 1) into the trained hierarchical recognition classifier obtained in step 3) for testing and verification to obtain the final hierarchical recognition classifier;
[0093] 5) Input the real-time acquired GNSS interference signal into the final hierarchical identification classifier obtained in step 4), and the final hierarchical identification classifier outputs the category of the GNSS interference signal.
[0094] The effectiveness of this invention can be further verified through the following experiments.
[0095] 1. Experimental conditions:
[0096] The hardware platform used in this invention experiment was: an Intel Core i7-3.20GHz processor and 32GB of memory.
[0097] The software platform used in this invention experiment was: Windows 10 operating system, Matlab R2022a, and Python 3.7.
[0098] 2. Experimental content and results analysis:
[0099] This invention uses a training set to train a hierarchical recognition classifier and a test set to test the performance of the hierarchical recognition classifier.
[0100] Table 1 shows the identification results of various interference types when the interference power is -110dBm, using the confusion matrix.
[0101] By calculating the interference recognition accuracy of the hierarchical recognition classifier under different interference power, the recognition accuracy variation curve of the method of the present invention under different interference power is obtained, as shown in Figure 4(a).
[0102] By calculating the false alarm rate of the hierarchical identification classifier under different interference power, the change curve of the false alarm rate of the method of the present invention under different interference power is obtained, as shown in Figure 4(b).
[0103] By calculating the recognition accuracy of the hierarchical recognition classifier under different training dataset sizes, the recognition accuracy variation curves of the method of this invention under different training dataset sizes are obtained, as shown below. Figure 5 As shown.
[0104] The following is a reference to Table 1 and Figure 4. Figure 5 The experimental results diagram further describes the performance of the present invention.
[0105] The labels STI, MTI, LCI, SCI, Pulse, NFMI, Spoofing, BLGNI, and NIF in Table 1 correspond to the following nine types of GNSS interference signals: single-tone interference, multi-tone interference, linear chirp interference, sinusoidal chirp interference, impulse interference, narrowband noise FM interference, spoofing interference, band-limited Gaussian noise interference, and no interference. Each column of the confusion matrix represents the predicted category of the GNSS interference signal, and each row represents the true category of the GNSS interference signal.
[0106] As shown in Table 1, even at an interference power of -110 dBm, the method of this invention achieves an accuracy rate of over 90% for identifying the eight types of interference, excluding pulse interference, demonstrating a high level of accuracy. However, 25.1% of the pulse interference was incorrectly identified as no interference. This is because the pulse interference repetition period defined in this invention is within the range of [0.02, 0.2] ms, while the GNSS interference signal acquisition time is 0.01 ms. It is possible that the acquired GNSS interference signal does not contain any pulse pairs.
[0107] In Figure 4(a), the horizontal axis represents different interference signal powers, and the vertical axis corresponds to the recognition accuracy corresponding to different interference powers.
[0108] As shown in Figure 4(a), even with interference signal power as low as -110 dBm, the average recognition accuracy of the method of this invention reaches 93.29%, indicating that the network has good performance in recognizing low-power GNSS interference. Furthermore, the average recognition accuracy gradually increases with increasing interference power.
[0109] In Figure 4(b), the horizontal axis represents different interference signal powers, and the vertical axis corresponds to the false alarm rate for different interference powers.
[0110] As can be seen from Figure 4(b), when the interference power is greater than -108dBm, the false alarm rate of the method of the present invention remains at around 0.01%, and even when the interference power is as low as -110dBm, the false alarm rate is only 3.31%.
[0111] Figure 5 The horizontal axis represents different numbers of training set samples, and the vertical axis corresponds to the recognition accuracy corresponding to different numbers of training set samples.
[0112] from Figure 5 As can be seen, the recognition accuracy of the method of the present invention is closely related to the number of training samples; the more training samples, the higher the recognition accuracy. Furthermore, even with only 400 training samples, the recognition accuracy is 85.62%, indicating that the method of the present invention can achieve high recognition accuracy without requiring excessive training data, saving a significant amount of time in data preparation and classifier training.
[0113] Table 1
[0114] STI MTI LCI SCI Pulse NFMI Spoofing BLGNI NIF STI 97.9％ 0.0％ 0.0％ 0.0％ 6.5％ 0.0％ 0.0％ 0.0％ 0.0％ MTI 0.0％ 100.0％ 0.0％ 0.0％ 0.0％ 0.0％ 0.0％ 0.0％ 0.0％ LCI 0.0％ 0.0％ 98.4％ 3.1％ 0.0％ 0.0％ 0.5％ 0.0％ 0.0％ SCI 0.0％ 0.0％ 0.0％ 94.3％ 0.0％ 0.0％ 0.0％ 0.0％ 0.0％ Pulse 0.0％ 0.0％ 0.0％ 0.0％ 68.4％ 0.0％ 0.0％ 0.0％ 0.0％ NFMI 0.0％ 0.0％ 1.6％ 0.0％ 0.0％ 90.5％ 9.4％ 0.0％ 0.0％ Spoofing 0.0％ 0.0％ 0.0％ 0.0％ 0.0％ 9.5％ 90.1％ 0.0％ 0.0％ BLGNI 0.0％ 0.0％ 0.0％ 0.0％ 0.0％ 0.0％ 0.0％ 100.0％ 0.0％ NIF 2.1％ 0.0％ 0.0％ 2.6％ 25.1％ 0.0％ 0.0％ 0.0％ 100.0％
Claims
1. A low-power GNSS interference identification method based on convolutional neural networks, characterized in that: The low-power GNSS interference identification method based on convolutional neural networks includes the following steps performed in sequence: 1) Generate 9 types of GNSS interference signals and construct an interference signal dataset, then divide the interference signal dataset into a training set and a test set according to the proportion; 2) Construct a hierarchical recognition classifier consisting of a time-frequency image classifier, an autocorrelation function image classifier, and a spectral flatness classifier; 3) Use the training set obtained in step 1) to train the hierarchical recognition classifier constructed in step 2), and obtain the trained hierarchical recognition classifier; the specific method is as follows: Input the training set obtained in step 1) into the hierarchical recognition classifier constructed in step 2), train the hierarchical recognition classifier, set the initial learning rate to 0.0001, select the Adam algorithm as the optimization algorithm, and select the cross-entropy loss function as the loss function; First, the GNSS interference signals in the training set are transformed using the short-time Fourier transform through a time-frequency image classifier. Then, the time-frequency transformation results of the GNSS interference signals are denoised to obtain the denoised time-frequency image of the GNSS interference signals as the first layer of features. If the interference is single-tone, multi-tone, linear chirp, sinusoidal chirp, or impulse, the time-frequency image classifier will directly output the recognition result. If the interference is either deception or narrowband noise FM, the autocorrelation function of the GNSS interference signal is calculated by an autocorrelation function image classifier, and an autocorrelation function image is plotted as the second layer of features. The autocorrelation function image is then used to perform secondary identification of the GNSS interference signal, thereby distinguishing between deception interference and narrowband noise FM interference. If the interference is band-limited Gaussian noise or no interference, the GNSS interference signal is subjected to fast Fourier transform by a spectral flatness classifier to obtain the spectral flatness of the GNSS interference signal as the third layer feature. The spectral flatness is then compared with a set threshold. If it is greater than the threshold, it is no interference; if it is less than the threshold, it is band-limited Gaussian noise interference. This process is used to identify the GNSS interference signal a second time. The trained hierarchical recognition classifier is obtained when the rate of change of the cross-entropy loss function is less than 0.1%. 4) Input the test set obtained in step 1) into the trained hierarchical recognition classifier obtained in step 3) for testing and verification to obtain the final hierarchical recognition classifier; 5) Input the real-time acquired GNSS interference signal into the final hierarchical identification classifier obtained in step 4), and the final hierarchical identification classifier outputs the category of the GNSS interference signal.
2. The low-power GNSS interference identification method based on convolutional neural networks according to claim 1, characterized in that: In step 1), the nine types of GNSS interference signals are single-tone interference, multi-tone interference, linear chirp interference, sinusoidal chirp interference, pulse interference, narrowband noise FM interference, deception interference, band-limited Gaussian noise interference, and no interference. The interference signal dataset is composed of all GNSS interference signals. Then, the interference signal dataset is divided into a training set and a test set at a ratio of 10:
1. The power of each GNSS interference signal is at Within the dBm range, the carrier initial phase is Within the range, other parameters of the interference signal are set as follows: the single-tone interference carrier frequency is... In the MHz range; multi-tone interference includes 2 single-tone interferences, each with a carrier frequency of... Within the MHz range; The sweep bandwidth of linear chirp interference and sinusoidal chirp interference is in Within the MHz range, the sweep period is Within the range; The bandwidth of narrowband noise FM interference is Within the MHz range; the bandwidth of band-limited Gaussian noise interference is Within the MHz range; the pulse interference repetition period is in Within the ms range; the Doppler frequency shift of deception interference is in Within the MHz range.
3. The low-power GNSS interference identification method based on convolutional neural networks according to claim 1, characterized in that: In step 2), the time-frequency image classifier includes: a first convolutional layer, a second convolutional layer, a first max-pooling layer, a third convolutional layer, a second max-pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a softmax layer; a channel attention mechanism is used after the second max-pooling layer; a residual connection module is used between the first and second convolutional layers; and a residual connection module is used between the first and third max-pooling layers; the kernel filter size of all convolutional layers and max-pooling layers is set to... The number of kernel filters in the first and second convolutional layers is 32, the number of kernel filters in the third convolutional layer is 64, the stride of the kernel filters in the convolutional layers is 1, the stride of the kernel filters in the max pooling layers is 2, the activation function of the convolutional layers is set to the ReLU activation function, and the number of neurons in the first, second and third fully connected layers is set to 512, 256 and 7 respectively. The autocorrelation function image classifier includes: a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a softmax layer; the kernel filter size of all convolutional layers and max-pooling layers is set to... The first convolutional layer has 32 kernel filters, the second convolutional layer has 64 kernel filters, the stride of the convolutional layer kernel filters is 1, the stride of the max pooling layer kernel filters is 2, and the number of neurons in the first, second, and third fully connected layers is set to 512, 256, and 2, respectively.
4. The low-power GNSS interference identification method based on convolutional neural networks according to claim 1, characterized in that: The short-time Fourier transform formula is as follows: ； in, This is a GNSS interference signal. For time variables, Therefore A time-centric window function , For discrete frequency points, , The imaginary unit, The result is the time-frequency transformation of the GNSS interference signal. For the number of points in the Fast Fourier Transform, The number of times the window function slides. .
5. The low-power GNSS interference identification method based on convolutional neural networks according to claim 1, characterized in that: The method for denoising the time-frequency transformation results of GNSS interference signals is as follows: The time-frequency transformation results of the calculated GNSS interference signal The time-frequency function value corresponding to its third quartile is selected as the threshold. Then, the time-frequency transformation result of the GNSS interference signal is denoised using the following soft threshold denoising method: ； in, For symbolic functions, it is expressed as follows: 。 6. The low-power GNSS interference identification method based on convolutional neural networks according to claim 1, characterized in that: The autocorrelation function is obtained by the following formula: ； in, For time delay, GNSS interference signal The length.
7. The low-power GNSS interference identification method based on convolutional neural networks according to claim 1, characterized in that: The spectral flatness is obtained by the following formula: ； in, For discrete frequency points, GNSS interference signal The result after performing a Fast Fourier Transform. The number of discrete frequency points.
8. The low-power GNSS interference identification method based on convolutional neural networks according to claim 1, characterized in that: In step 2), the specific operation methods of the channel attention mechanism and residual connection module in the time-frequency image classifier are as follows: The channel attention mechanism, through squeezing and activation operations, enables the network to adaptively learn the weights of each channel and adjust the importance of each channel's features in subsequent network layers, thereby improving the classifier's performance. The input to the squeezing operation is a variable of size [value missing]. Feature map , The number of channels in the input feature map, for the feature map Perform global average pooling to compress it into a size of eigenvectors ; Eigenvector The first in element Represented as: ； in, For the first Feature maps of each input channel These are the length and width of the feature map for that channel, respectively. In the excitation operation, the feature vector generated by the squeeze operation is... The feature maps of the input are learned and generated through fully connected layers and non-linear activation functions. weight vector Weight vector The calculation formula is as follows: ； in, , , Less than ; for Activation function for The activation function will eventually learn the weight vector. Applied to the input feature map For each channel, the final output feature map is obtained. ,in For the first The weighted result of the feature maps of each input channel is represented as follows: ； The residual connection module skips some network layers from the input features and connects them directly to the output.