GNSS interference signal detection method based on ResNet50-CBAM network
Through the GNSS interference signal detection method of the ResNet50-CBAM network, combined with the convolutional attention module CBAM, the recognition error rate problem caused by the increase in the number of models in the prior art is solved, and efficient identification of multiple interference signals is achieved, especially Excellent in low power signal detection.
Patent Information
- Application Number
- CN202410982420.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-08
AI Technical Summary
Existing satellite navigation interference signal detection and classification methods will lead to gradient disappearance and degradation when the number of model layers increases, resulting in an increase in recognition error rate, making it difficult to effectively identify low-power interference signals.
The GNSS interference signal detection method based on the ResNet50-CBAM network is adopted, and the time-frequency analysis and feature extraction are performed through the ResNet50 residual network model combined with the convolution attention module CBAM to improve the recognition ability of different interference signal types.
It improves the recognition and detection efficiency of various types of interference signals, especially in the detection of low-power interference signals, enhances the model's ability to express image features and reduces the recognition error rate.
Smart Images

Figure CN118915097A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of satellite navigation interference suppression, and in particular to a GNSS interference signal detection method based on a ResNet50-CBAM network. Background Art
[0002] In recent years, the satellite navigation system GNSS has been widely used in the fields of transportation, agriculture, forestry, fishery, hydrological monitoring, meteorological observation, communication timing, power dispatching, disaster relief and reduction, public safety, etc., serving the country's important infrastructure and generating significant economic and social benefits. Satellite navigation signals are greatly affected when they pass through the atmosphere and ionosphere after long-distance transmission. The power when they reach the ground receiver is very weak, usually around -160dBW, and are easily affected by high-power intentional noise signals (the so-called interference) or even unintentional noise generated on the surface of the earth. Therefore, the current improvement of anti-interference technology for satellite navigation receivers needs to be solved urgently in order to better deal with various complex interference situations. In order to avoid the influence of various interferences on navigation signals and improve the anti-interference performance of navigation receivers, it is necessary to classify the interference signals at the front end of the receiver so as to select the corresponding anti-interference algorithm according to the different interference signals. The work of effectively characterizing and identifying interference types using deep learning plays a vital role in designing the anti-interference algorithm capabilities of any receiver. Existing satellite navigation interference signal detection and classification methods mainly use one or more combination methods such as artificial neural network (ANN), recursive neural network (RNN), support vector machine (SVM) and convolutional neural network (CNN) to classify different interference signals. The recognition performance of the convolutional neural network of the above algorithm will gradually improve with the increase of its layers, but the increase in the number of model layers will lead to problems such as gradient disappearance and degradation, making model training difficult, increasing the network recognition error rate, and causing the performance of the neural network to degrade. Summary of the invention
[0003] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and provide a GNSS interference signal detection method based on the ResNet50-CBAM network to realize the detection and classification of satellite navigation suppression interference.
[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is: a GNSS interference signal detection method based on the ResNet50-CBAM network, comprising the following steps:
[0005] Step 1: Generate an original satellite navigation signal using a satellite navigation signal simulation source, generate an interference signal using a vector signal generator, and add the interference signal to the satellite navigation signal via a combiner;
[0006] The interference signal is divided into sweep frequency interference, continuous wave interference, pulse interference, frequency hopping interference, frequency modulation interference and narrowband interference;
[0007] Step 2: Perform time-frequency analysis on the satellite navigation signal mixed with the interference signal, and store the time-frequency analysis results as an image;
[0008] Firstly, short-time Fourier transform or Wigner-Ville transform is applied to the satellite navigation signal mixed with the interference signal for time-frequency analysis. The power spectrum of the original satellite navigation signal is calculated by the Welch method and stored as an image.
[0009] Step 3: Build and train a satellite navigation interference signal detection model to detect interference signals;
[0010] The satellite navigation interference signal detection model adopts the ResNet residual network model, including a convolution layer, a pooling layer, a fully connected layer and a residual module; the ResNet residual network model adopts the ResNet50 network model structure as the backbone feature extraction network, including 1 convolution block, 4 residual blocks and 1 output block; wherein the convolution block includes a convolution layer with a size of 7x7, a batch normalization operation layer BN, an activation layer and a 3x3 maximum pooling layer; the residual block includes an identical residual block and a convolution residual block; the number of convolution residual blocks in each residual block is the same while the number of identical residual blocks is different; wherein the identical residual block and the convolution residual block are both composed of bottleneck residual units with sizes of 1x1, 3x3, and 1x1; the output of the bottleneck residual unit is composed of multiple cascaded convolution layer inputs and cascaded convolution outputs, the cascaded convolution layer input is called an identical mapping, and the cascaded convolution output is called a residual mapping;
[0011] After the third convolutional layer of the bottleneck residual unit, a multi-frequency channel attention network FcaNet is used to improve the bottleneck residual unit and refine the output features of the residual structure to obtain better classification results.
[0012] The input of the ResNet50 network model is a satellite navigation signal image after time-frequency transformation; it is standardized, the mean is subtracted and divided by the standard deviation; the initial convolution layer uses a 7x7 convolution kernel, a step size of 2, and a channel number of 64; then a 3x3 maximum pooling is performed with a step size of 2 to reduce the size of the feature map; each stage contains multiple residual blocks, each residual block consists of 3 convolution layers, and the convolution kernels of each convolution layer are 1x1, 3x3 and 1x1, respectively, for extracting features of different scales; in each residual block, the input is directly added to the output through a residual connection to form residual learning; the output of the last residual block is globally average pooled to convert the feature map into a vector of a fixed size; the fully connected layer is used to map the output of the global average pooling to the final classification result; the last layer uses a Softmax activation function to convert the output of the network into a category probability distribution;
[0013] In order to improve the feature extraction and expression capabilities of the ResNet50 network, the ResNet residual network model introduces a convolutional attention module CBAM on the basis of the backbone feature extraction network ResNet50 to form a ResNet50-CBAM feature extraction network; the ResNet50-CBAM network suppresses invalid feature information by enhancing the spatial and channel information aggregation capabilities of channel information; the channel attention module in the convolutional attention module enables the network to selectively emphasize important channel information by learning the attention weights of each channel of the input feature map; the spatial attention module learns the positional relationships of the input feature map under the same channel to adjust the attention weights, which helps the network to focus on the target area more effectively;
[0014] CBAM is integrated into the convolutional residual blocks of ResNet50. The feature map output by the previous residual block is used as the input feature matrix F of CBAM after convolution. After being processed by the channel attention module, the feature matrix F' is obtained, and the new feature matrix F' is obtained by dot multiplication with the input feature map F; then, F' is dot multiplied with the feature matrix F" processed by the spatial attention module to obtain the final feature matrix F"'; finally, F"' is superimposed with the input feature map as the input of the next residual block; CBAM is organically integrated into the ResNet50 network, effectively improving the network's ability to learn and express feature information of different types of interference signals.
[0015] Step 4: The satellite navigation signal mixed with the interference signal is subjected to short-time Fourier transform or Wigner-Ville transform to obtain an image corresponding to the interference signal. The satellite navigation interference signal detection model determines the type of the interference signal based on the transformed image and the features in the time domain and frequency domain.
[0016] The beneficial effects of adopting the above technical solution are: the GNSS interference signal detection method based on the ResNet50-CBAM network provided by the present invention is
[0017] (1) A satellite navigation interference signal detection model is constructed and trained to detect satellite navigation interference signals. The CBAM attention mechanism can adaptively adjust the importance of different channels and spatial positions, thereby improving the model's ability to express image features, enabling the model to better capture the feature information in the interference signal and improve the recognition and detection efficiency of various types of interference signals.
[0018] (2) Due to the fragility of satellite navigation signals, even weak interference signals can affect the performance of GNSS receivers. The existing detection methods are easy to detect interference signals with high power, but are less effective in detecting low-power interference. The present invention achieves low-power interference detection by extracting statistical features and obtaining further details of the original data. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 A flowchart of a GNSS interference signal detection method based on a ResNet50-CBAM network provided in an embodiment of the present invention;
[0020] Figure 2 A Welch transformation calculation flow chart provided in an embodiment of the present invention;
[0021] Figure 3 A structural diagram of the ResNet50 network model provided in an embodiment of the present invention;
[0022] Figure 4 A structural diagram of an improved bottleneck residual unit provided in an embodiment of the present invention;
[0023] Figure 5 This is a structural diagram of the ResNet50-CBAM network model provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0024] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.
[0025] In this embodiment, a GNSS interference signal detection method based on a ResNet50-CBAM network is as follows: Figure 1 As shown, the following steps are included:
[0026] Step 1: Generate an original satellite navigation signal using a satellite navigation signal simulation source, generate an interference signal using a vector signal generator, and add the interference signal to the satellite navigation signal via a combiner;
[0027] The interference signal is divided into sweep frequency interference, continuous wave interference, pulse interference, frequency hopping interference, frequency modulation interference and narrowband interference;
[0028] Different types of interference signal models are:
[0029] (1) Sweep frequency interference is shown in the following formula:
[0030]
[0031] Among them, a(t) and φ(t) are the amplitude and phase of the swept frequency interference respectively, f(t) is the instantaneous frequency, and for the linear swept frequency interference model, f(t) = f0 + kt, where the frequency changes linearly with time, and t represents time;
[0032] The characteristic of swept frequency jammers is that they can generate a suppressive jammer signal whose carrier frequency varies with the GNSS signal band. The characteristics of swept frequency jammer signals can be classified by shape, sweep range, sweep rate and power.
[0033] (2) Continuous wave interference
[0034] Continuous wave interference (CWI) consists of a pure sine wave of constant amplitude and frequency, which appears as a single tone in the frequency domain, as shown in the following formula:
[0035] J(t)=Aexp(j2πf i t+φ i )
[0036] Where A is the amplitude, f i and φ i are the frequency and phase of the continuous wave interference signal respectively.
[0037] (3) Pulse interference
[0038]
[0039] Where P is the interference power, α is the pulse width, Δt is the pulse time interval, t k is the emission time of the kth pulse pair.
[0040] (4) Frequency Hopping Interference
[0041] Frequency hopping interference introduces discontinuities in the instantaneous frequency by repeatedly switching the carrier frequency.
[0042]
[0043] Among them, f m (t) is a function that randomly creates a set of m consonants occupying a specific frequency range, A is the amplitude, and W is the number of random generation trials.
[0044] (5) FM interference
[0045] FM jamming is a form of angle modulation in which the instantaneous frequency of the carrier varies linearly, which can be used by jammers. FM jamming of a sinusoidally modulated signal can be expressed as follows:
[0046] J(t)=Eexp(j2πf n t+βsin(2πf m t))
[0047] Where E is the carrier amplitude, f n is the interference frequency, f m is the frequency of the modulating signal, β represents the modulation index;
[0048] (6) Narrowband interference
[0049] Narrowband interference (NBI) or jammer is interference whose spectrum occupancy is less than the GNSS signal bandwidth. NBI can be called narrowband Gaussian, narrowband phase or narrowband swept frequency interference, as shown in the following formula:
[0050]
[0051] Where β is the modulation index, n(λ) represents a stationary random process with zero mean;
[0052] Basically, the narrowband interference signal presented is a Gaussian White Noise (WGN) filtered with a rectangular pulse shape and 2Mhz bandwidth.
[0053] GNSS signals come from the GPS constellation, belong to the L1 frequency band, and have a center frequency of 1575.42MHz. The carrier-to-noise ratio (C / N0) of the satellite navigation signal is between 25 and 50 dBHz. Within the range of 25 to 50 dBHz, the signal quality ranges from poor reception to good reception, and may be interfered with in different ways. In this embodiment, the satellite navigation signal is simulated using the NSS8900 navigation simulator, and the interference signal is generated by the SMBV100A vector signal / interference generator.
[0054] To simulate the generation of interference signals, the GNSS signal was superimposed with different types of interference signals through a combiner. The received signal was filtered by a wideband filter with a bandwidth of 20MHz and the sampling frequency fs was set to 40MHz. The above results were obtained considering a large number of situations caused by the combination of GNSS signal carrier-to-noise ratio (C / N0) values in 1dBHz steps between 25 and 50dBHz. Different power interference levels between -142dBW and -107dBW. In this way, different conditions of the interference signal ratio were taken into account in the results. In the Monte Carlo simulation, different situations of the initial relative phase of the interference and GNSS signals were simulated.
[0055] Step 2: Perform time-frequency analysis on the satellite navigation signal mixed with the interference signal, and store the time-frequency analysis results as an image;
[0056] Firstly, the short-time Fourier transform (STFT) or Wigner-Ville transform (WVT) is applied to the satellite navigation signal mixed with the interference signal for time-frequency analysis, and the time-frequency analysis result (Time-Frequency Representation, TFR) is processed by the Welch method and stored as an image;
[0057] The short-time Fourier transform (STFT) method is used for satellite navigation signals mixed with interference signals. Figure 2 As shown, the details are as follows:
[0058] Using a fixed-length window function, the non-stationary satellite navigation signal is split into shorter segments (frames) of N samples, and the Fourier transform of each frame is evaluated separately; the STFT can be written as:
[0059]
[0060] Among them, m is the translation parameter, ω N [nm] is the window function, x[n] is the original satellite navigation signal, n is the time variable, and ω is the frequency variable;
[0061] The Wigner-Ville transform (WVT) method for satellite navigation signals mixed with interference signals is as follows:
[0062] For a satellite navigation signal x[n] considering a window of N samples, the Wigner-Ville transform can be written as:
[0063]
[0064] Among them, x * represents the conjugate of x, n and m represent the index numbers of the time vector and frequency vector respectively;
[0065] In order to avoid the cross-term effect of the quadratic form, the smoothed pseudo WVT (SPWVT) is used to analyze the satellite navigation signal. SPWVT uses independent window functions g(n) and h(m) to smooth in time and frequency, respectively:
[0066]
[0067] Among them, g(n) is the time smoothing window function, h(m) is the lag window function; WVT provides high-resolution time and frequency representation for non-stationary signals. In addition, it has the advantages of satisfying instantaneous power in time, energy spectrum in frequency, and total energy in time and frequency planes.
[0068] In order to extract the features in the frequency domain, the Welch method is used to calculate the power spectrum of the original satellite navigation signal. The Welch power spectrum is a power spectrum formed by segmenting and windowing the data. The calculation process is as follows: Figure 3 To calculate the Welch power spectrum, we need to determine the length of the converted signal, the step size of the window function each time, and the window function category. The window function is selected as a rectangular window function, and the formula is:
[0069]
[0070] Among them, W m (b) is the mth window function at point b; d is the interception length of the window function; e is the interception position of the window function each time; m is the number of interceptions.
[0071] The interception formula is:
[0072] g m (b) = g[b]*W m (b)
[0073] Among them, g m (b) is the data after the window function is intercepted for the mth time; g[b] is the intercepted original satellite navigation signal data;
[0074] After determining the above information, calculate the Welch power spectrum of the satellite navigation signal as shown in the following formula:
[0075]
[0076] Among them, P welch is the Welch power spectrum of the satellite navigation signal, M is the maximum number of intercepts, FFT is the fast Fourier transform, and conj is the conjugate of the complex number;
[0077] In this example, the image (TFR) is obtained by STFT and WVT with an interval of 100 μs, which is sufficient to characterize the different behaviors of the jammer in time. Table 1 shows the STFT parameters used to create the spectrogram.
[0078] Table 1 STFT parameters
[0079]
[0080]
[0081] Statistical features in the time and frequency domains are extracted within a 200μs window of the input signal. Two datasets are generated, each representing a different signal processing technique (STFT and WVT). Each dataset consists of 17 classes, one of which refers to GNSS signals without interference, and the other 16 classes refer to GNSS signals with various interferences. The dataset contains 111,384 samples, with an equal number of samples in each class. The C / N0 of the GNSS signal is 40dBHz and the interference power is -120dBW.
[0082] Step 3: Build and train a satellite navigation interference signal detection model to detect interference signals;
[0083] The satellite navigation interference signal detection model adopts the ResNet residual network model, including a convolutional layer, a pooling layer, a fully connected layer and a residual module; as the convolutional neural network goes deeper, the accuracy tends to saturation, the training complexity increases, and the convolutional neural network tends to degenerate. The ResNet model skips some layers of the model and directly connects the input of one layer to other layers, which is called a skip connection.
[0084] The ResNet residual network model uses the ResNet50 network model structure as the backbone feature extraction network, such as Figure 3 As shown, it includes 1 convolution block, 4 residual blocks and 1 output block; wherein the convolution block includes a convolution layer of size 7x7, a batch normalization operation layer BN (Batch Normalization), an activation layer (GeLU function) and a 3x3 maximum pooling layer; the residual block includes an identical residual block (i.e., a residual structure that does not change the size) and a convolution residual block (i.e., a residual structure that adds a scale); the number of convolution residual blocks in each residual block is the same, while the number of identical residual blocks is different; in this embodiment, the number of convolution residual blocks is 1, and the number of identical residual blocks is 2, 3, 5, and 2, respectively. The identical residual block and the convolution residual block are both composed of bottleneck residual units of sizes 1x1, 3x3, and 1x1; Figure 4It is the input-output relationship of a bottleneck residual unit in the network. The output of the bottleneck residual unit is composed of multiple cascaded convolution layer inputs and cascaded convolution outputs. The cascaded convolution layer input is called identity mapping, and the cascaded convolution output is called residual mapping.
[0085] Figure 4 In the figure, x and F(x) are the input and output of the bottleneck residual unit respectively; H(x) is the output after the cascade convolution through the residual unit input. The output of the residual unit is
[0086] H(x)=F(x)+x
[0087] When the network model's learning ability reaches the optimal level, if the network continues to be deepened for training, F(x) will be forced to be trained to 0, leaving only the identity mapping H(x) = x, so that the network will always be in the best state.
[0088] The ResNet50 network model performs well in terms of depth, training time, and accuracy. The ResNet50 network model has 23 million parameters and 49 convolutional layers, the last of which is connected to an average pooling layer and enters a fully connected layer to perform classification tasks.
[0089] The input of the ResNet50 network model is a satellite navigation signal image after time-frequency transformation, with a size of 224x224x3; normalization is performed, the mean is subtracted and divided by the standard deviation, so that the network can better process image data; the initial convolution layer uses a 7x7 convolution kernel, a step size of 2, and a channel number of 64; then a 3x3 maximum pooling is performed with a step size of 2 to reduce the size of the feature map. Each stage contains multiple residual blocks, each of which consists of 3 convolution layers, and the convolution kernels of each convolution layer are 1x1, 3x3 and 1x1, respectively, for extracting features of different scales; in each residual block, the input is directly added to the output through the residual connection to form residual learning; the output of the last residual block is globally averaged pooled to convert the feature map into a vector of fixed size, which reduces the number of parameters and helps to reduce the risk of overfitting; the fully connected layer is used to map the output of the global average pooling to the final classification result; the last layer uses the Softmax activation function to convert the output of the network into a category probability distribution.
[0090] In order to improve the feature extraction and expression capabilities of the ResNet50 network, the ResNet residual network model introduces a convolutional attention module (CBAM) based on the backbone feature extraction network ResNet50 to form a ResNet50-CBAM feature extraction network; the ResNet50-CBAM network suppresses invalid feature information by enhancing the spatial and channel information aggregation capabilities of channel information; among them, the channel attention module in the convolutional attention module learns the attention weights of each channel of the input feature map, so that the network can selectively emphasize important channel information; the spatial attention module learns the various positional relationships of the input feature map under the same channel to adjust the attention weights, which helps the network to focus on the target area more effectively.
[0091] Since the input and output feature matrices of CBAM have the same dimensions, they can be inserted into the ResNet50 model to effectively learn the feature information of the interference signal type. CBAM is integrated between the convolutional residual blocks of ResNet50. The specific structure is as follows: Figure 5 As shown. The feature map output by the previous residual block is used as the input feature matrix F of CBAM after convolution. After being processed by the channel attention module, the feature matrix F' is obtained, and the new feature matrix F' is obtained by dot multiplication with the input feature map. Next, F' is dot-multiplied with the feature matrix F" processed by the spatial attention module to obtain the final feature matrix F". Finally, F" is superimposed with the input feature map as the input of the next residual block. CBAM is organically integrated into the ResNet50 network, effectively improving the network's ability to learn and express feature information of different types of interference signals.
[0092] When calculating channel attention, each channel needs a learnable scalar value to calculate the attention weight. The scalar calculation function generally uses global average pooling (GAP). However, the simple de-averaging method of GAP discards a lot of information and cannot fully capture the diversity of each channel. Since GAP is a special form of discrete cosine transform (DCT), channel attention is extended to the frequency domain based on this. After the third convolutional layer of the bottleneck residual unit, the multi-frequency channel attention network FcaNet is used to improve the bottleneck residual unit and refine the output features of the residual structure to obtain better classification results. The FcaNet network obtains the weights corresponding to different feature channels through learning, and then improves the expression ability of useful features while suppressing the expression ability of useless features according to the size of the weights. In the FcaNet network, unlike the global average pooling (GAP) method that only uses the lowest frequency band, the network innovatively fuses multi-band components and significantly expands the dimension of feature extraction. By fusing information from different frequency bands to enrich the diversity of feature sets, the network's ability to extract target features of different interference signals is enhanced, and the recognition accuracy is improved.
[0093] In this embodiment, the training process for the satellite navigation interference signal detection model is:
[0094] Split the dataset: The image dataset is split into three subsets, and the training process uses 50% of the dataset. 15% of the dataset is used to tune the parameters and find the best model (validation dataset), while 35% is used for unbiased evaluation of the final model. To split the dataset, a random state function controls the shuffling process, which means that all samples in the training, validation, and test datasets are the same for all models.
[0095] Optimization: The training process uses the Adam optimization algorithm to update the neural network parameters, the batch size is 64 samples, and the initial learning rate is 0.001. Adaptive learning rate is considered. In the model, the entire training data is trained once, which is called an epoch. If the performance does not improve with the increase of the current learning rate after two epochs, the learning rate is divided into two. The loss function used is categorical cross entropy (softmax loss).
[0096] Training epochs: The number of epochs for training is set to 15, and all samples in the training dataset are passed through the model 15 times. In addition, the early stopping (3 epochs) method is defined to stop training once the performance stops improving on the validation dataset.
[0097] The evaluation metrics used to measure the performance of the method for the ResNet-50 network model are:
[0098] Confusion Matrix: Displays the number of correct and incorrect predictions for each class and counts them to provide a basis for the results of the classification task. The confusion matrix demonstrates how the classification model makes predictions when confused. Table 2 shows the confusion matrix for the binary classification results.
[0099] Table 2 Confusion matrix of binary classification results
[0100]
[0101] A True Positive (TP) is a correct positive prediction, i.e. the model correctly identifies a positive class, the predicted signal is a positive example, and the true signal label is also a positive example. A False Negative (FN) is an incorrect negative prediction, i.e. the model cannot identify the positive class, the predicted signal is a negative example, and the true signal label is a positive example. A False Positive (FP) is an incorrect positive prediction, i.e. the model identifies a positive class, but it is actually a negative class, the predicted signal is a positive example, and the true signal label is a negative example. A True Negative (TN) is a correct negative prediction, i.e. the model correctly identifies a negative class, the predicted signal is a negative example, and the true signal label is also a negative example.
[0102] Overall accuracy: Overall accuracy is determined by calculating the total number of correctly classified samples divided by the total number of samples. It can be expressed as:
[0103]
[0104] Where n c is the category, n s is the total number of samples, TP i The total number of samples correctly classified for the i-th class.
[0105] F1-score: The goal of F1-score is to create a single metric by combining the precision p and recall r criteria, defined as the harmonic mean of precision and recall:
[0106]
[0107] Among them, precision refers to the proportion of correctly predicted positive samples (including true positives and false positives) to the total number of positive prediction samples (p=TP / (TP+FP)), and recall refers to the proportion of correctly predicted positive samples to correctly classified samples (including true positives and true negatives) (r=TP / (TP+TN)).
[0108] Inference time: The time required for the model to apply a trained neural network to new data is called inference time or network latency. This metric is a key indicator for deploying network models in practical applications.
[0109] Task completion time: Task completion time refers to the complete process from interference signal input to detection and classification (such as Figure 1 The time required for the model to be processed includes the signal preprocessing time and the model detection and classification time.
[0110] Step 4: The satellite navigation signal mixed with the interference signal is subjected to short-time Fourier transform (STFT) or Wigner-Ville transform (WVT) to obtain the image corresponding to the interference signal. The satellite navigation interference signal detection model determines the type of the interference signal based on the transformed image and the features in the time domain and frequency domain.
[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.
Claims
1. A GNSS interference signal detection method based on ResNet50-CBAM network, characterized by: The following steps are involved: Step 1: Generate an original satellite navigation signal using a satellite navigation signal simulation source, generate an interference signal using a vector signal generator, and add the interference signal to the satellite navigation signal via a combiner; Step 2: Perform time-frequency analysis on the satellite navigation signal mixed with the interference signal, and store the time-frequency analysis results as an image; Step 3: Build and train a satellite navigation interference signal detection model to detect interference signals; The satellite navigation interference signal detection model adopts the ResNet residual network model, including a convolution layer, a pooling layer, a fully connected layer and a residual module; the ResNet residual network model adopts the ResNet50 network model structure as the backbone feature extraction network, including 1 convolution block, 4 residual blocks and 1 output block; wherein the convolution block includes a convolution layer with a size of 7x7, a batch normalization operation layer BN, an activation layer and a 3x3 maximum pooling layer; the residual block includes an identical residual block and a convolution residual block; the number of convolution residual blocks in each residual block is the same but the number of identical residual blocks is different; wherein the identical residual block and the convolution residual block are both composed of bottleneck residual units with sizes of 1x1, 3x3, and 1x1; the output of the bottleneck residual unit is composed of multiple cascaded convolution layer inputs and cascaded convolution outputs, the cascaded convolution layer input is called an identical mapping, and the cascaded convolution output is called a residual mapping; Step 4: The satellite navigation signal mixed with the interference signal is subjected to short-time Fourier transform or Wigner-Ville transform to obtain an image corresponding to the interference signal. The satellite navigation interference signal detection model determines the type of the interference signal based on the transformed image and the features in the time domain and frequency domain.
2. According to a GNSS interference signal detection method based on ResNet50-CBAM network according to claim 1, it is characterized in that: The interference signals are classified into swept frequency interference, continuous wave interference, pulse interference, frequency hopping interference, frequency modulation interference and narrowband interference.
3. According to a GNSS interference signal detection method based on ResNet50-CBAM network according to claim 1, it is characterized in that: The step 2 applies short-time Fourier transform or Wigner-Ville transform to the satellite navigation signal mixed with the interference signal to perform time-frequency analysis, and calculates the power spectrum of the original satellite navigation signal through the Welch method based on the time-frequency analysis result and stores it as an image.
4. According to a GNSS interference signal detection method based on ResNet50-CBAM network according to claim 1, it is characterized in that: The bottleneck residual unit adopts a multi-frequency channel attention network FcaNet after the third convolutional layer to refine the output features of the residual structure to obtain better classification effect.
5. According to claim 4, a GNSS interference signal detection method based on ResNet50-CBAM network is characterized in that: The input of the ResNet50 network model is a satellite navigation signal image after time-frequency transformation; it is standardized, the mean is subtracted and divided by the standard deviation; the initial convolution layer uses a 7x7 convolution kernel, a step size of 2, and a channel number of 64; then a 3x3 maximum pooling is performed with a step size of 2 to reduce the size of the feature map; each stage contains multiple residual blocks, each residual block consists of 3 convolution layers, and the convolution kernels of each convolution layer are 1x1, 3x3 and 1x1, respectively, for extracting features of different scales; in each residual block, the input is directly added to the output through a residual connection to form residual learning; the output of the last residual block is globally average pooled to convert the feature map into a vector of a fixed size; the fully connected layer is used to map the output of the global average pooling to the final classification result; the last layer uses a Softmax activation function to convert the output of the network into a category probability distribution.
6. The GNSS interference signal detection method based on the ResNet50-CBAM network according to claim 5 is characterized in that: The ResNet residual network model introduces a convolutional attention module CBAM on the basis of the backbone feature extraction network ResNet50 to form a ResNet50-CBAM feature extraction network; the ResNet50-CBAM network suppresses invalid feature information by enhancing the spatial and channel information aggregation capabilities of channel information; wherein, the channel attention module in the convolutional attention module enables the network to selectively emphasize important channel information by learning the attention weights of each channel of the input feature map; the spatial attention module learns the positional relationships of the input feature map under the same channel to adjust the attention weights, which helps the network to focus on the target area more effectively; CBAM is integrated into the convolutional residual blocks of ResNet50. The feature map output by the previous residual block is used as the input feature matrix F of CBAM after convolution. After being processed by the channel attention module, the feature matrix F' is obtained, and the new feature matrix F' is obtained by dot multiplication with the input feature map F; then, F' is dot multiplied with the feature matrix F" processed by the spatial attention module to obtain the final feature matrix F"'; finally, F"' is superimposed with the input feature map as the input of the next residual block; CBAM is organically integrated into the ResNet50 network, effectively improving the network's ability to learn and express feature information of different types of interference signals.
Citation Information
Cited By
Identification method and device for satellite interference signal and storage medium
CN119439201A
Satellite navigation interference signal database and construction method
CN120180273A
Satellite navigation interference signal type identification method and system
CN120180307A
Low-power-consumption instantaneous frequency measurement system
CN120316488A
Method for taking provided GNSS-Relevant route information into account in the GNSS-based localization of vehicles
US12650522B2
