Radio signal waveform intelligent sensing method and system fusing multi-domain features

Abstract

The invention discloses a radio signal waveform intelligent sensing method and system fusing multi-domain features, and belongs to the field of wireless communication, and the method comprises the steps: carrying out the feature extraction of an input sequence, and extracting the multi-mode features of a signal through different modes; fusing the multi-mode features into fusion features with richer information through an attention mechanism; and finally, further classifying the fused features to obtain a final sensing signal modulation mode. According to the invention, the accuracy of target signal sensing can be improved.

Description

Technical Field

[0001] The present invention relates to the field of wireless communications, and more specifically, to a method and system for intelligently sensing radio signal waveforms integrating multi-domain features. Background Art

[0002] Targets are usually equipped with multiple devices such as communications and radars to ensure the execution of tasks. Depending on the task, the target will emit different radio signal waveforms, such as communication and radar terminals selecting different frequencies, bandwidths, modulation methods, etc. By sensing the target's radio signal, the target's working status can be analyzed to provide support for the judgment of the target's intentions.

[0003] Traditional radio signal waveform sensing technologies can be divided into two categories:

[0004] (1) Likelihood-based method. This method transforms the modulation perception problem of radio signal waveform into a multiple hypothesis testing problem, and identifies the modulation mode of the signal by setting multiple hypothesis thresholds and maximum likelihood ratio theory. The disadvantage of this method is that the discrimination threshold value is difficult to determine, the maximum likelihood ratio calculation is complex, and it is sensitive to the signal environment.

[0005] (2) Feature-based method. This method extracts multiple physical or statistical characteristics of the received signal and uses a classifier to perceive the modulation mode of the radio signal waveform. In practical applications, the selection of features is heavily dependent on manual work, and the performance is poor when identifying high-order modulated radio signals.

[0006] At present, data-driven deep learning technology has received widespread attention in the field of radio signal waveform perception. Existing advanced technologies include:

[0007] (1) Architecture based on Convolution Neural Network (CNN). CNN extracts modulation pattern features from radio signal waveforms by performing convolution operations on the sequence or transform domain image of the received signal. As the number of convolution layers increases, the receptive field of the output neuron also increases, and the representation ability of the output feature map is usually stronger.

[0008] (2) Architecture based on the Complex Value Convolution Neural Network (CVCNN). The CVCNN architecture usually processes complex sequences of signals. The principle is similar to that of the CNN architecture. CVCNN treats the input sequence as a complex number and extracts the interaction features between the real and imaginary parts of the sequence through convolution operations to perceive the modulation mode of the signal.

[0009] (3) Long Short-Term Memory (LSTM)-based architecture. LSTM architecture usually processes the time domain sequence of the signal and perceives the modulation mode of the signal through the time correlation of the input sequence.

[0010] The above methods all have certain limitations. Specifically, the features extracted by methods (1) and (2) are relatively simple. If only one of the methods is used, the radio signal waveform perception performance is often poor. Method (3) cannot achieve complete parallelism in structure due to the use of the time correlation of the input, which will result in high latency. Summary of the invention

[0011] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method and system for intelligently sensing radio signal waveforms that integrates multi-domain features, thereby improving the accuracy of target signal perception.

[0012] The object of the present invention is achieved through the following solutions:

[0013] A radio signal waveform intelligent perception method integrating multi-domain features, comprising:

[0014] Perform feature extraction on the input sequence and extract the multi-modal features of the signal in different ways; then fuse the multi-modal features into more informative fused features through the attention mechanism;

[0015] Finally, the fused features are further classified and processed to obtain the modulation mode of the final perception signal.

[0016] Furthermore, the input sequence includes IQ and FFT sequences.

[0017] Furthermore, the feature extraction of the input sequence extracts the multi-mode features of the signal in different ways, which specifically includes the sub-steps of performing real convolution RVC and complex convolution CVC operations on the two groups of sequences respectively, so as to obtain multiple groups of feature streams.

[0018] Furthermore, the method of fusing the multi-modal features into more informative fused features through the attention mechanism specifically includes the following sub-steps:

[0019] The convolution layer is split into two parts using the depthwise separable convolutional layer DSC, which consists of channel dimension convolution and 1×1 convolution. The transformation of the output feature map dimension is completed by 1×1 convolution.

[0020] In addition, the channel attention mechanism is used to correct the feature strength between channels by modeling the interdependence between channels. Specifically, the feature map is first compressed and the global average is calculated according to the channel dimension; then the weight corresponding to each channel of the feature map is obtained through the activation function, and finally the feature map is multiplied by the corresponding weight to obtain the fused feature.

[0021] Furthermore, the activation function is a Sigmoid activation function.

[0022] Furthermore, the method further includes a sub-step of providing a jump connection between the input and output of the attention mechanism to prevent the gradient from disappearing.

[0023] Furthermore, the fusion features are further classified and processed to obtain the modulation mode of the final perception signal, which specifically includes the following sub-steps: after receiving the fusion features, the fusion features are further processed using a residual module and a fully connected layer module, and mapped to the perception results.

[0024] Furthermore, after receiving the fused features, the fused features are further processed using a residual module and a fully connected layer module, and mapped to the perception results, specifically including the following sub-steps:

[0025] The residual module consists of two residual units and a pooling layer. Each residual unit consists of two RVC layers and a skip connection. The skip connection is represented as:

[0026] FF′=FF+f(FF),

[0027] Among them, FF represents the input feature map, f(FF) represents the feature map after two RVC operations, and FF′ represents the output feature map after the jump connection. During back propagation, even if the gradient of the convolution output f(FF) to the input feature map FF is 0, the jump connection output FF′ can ensure that the gradient to FF is at least 1, which is used to avoid the gradient disappearance phenomenon. The pooling layer is used to reduce the number of output neurons to prevent overfitting.

[0028] The fully connected layer module first flattens the feature map into a one-dimensional vector, and then inputs the vector into the fully connected layer, namely:

[0029] f k =f act (w k f k-1 )+b k ,k=1,2,…n,

[0030] Among them, n represents the number of fully connected layers, f k represents the output of the kth fully connected layer, w k and b kare the weight and bias of the kth fully connected layer respectively.

[0031] Furthermore, the feature extraction of the input sequence extracts the multi-mode features of the signal in different ways, specifically including: extracting the time domain and frequency domain features of the radio signal, and mapping them to form real and complex multi-mode features.

[0032] A radio signal waveform intelligent perception system integrating multi-domain features, comprising a feature extraction module, a feature fusion module and a classifier;

[0033] The feature extraction module is used to extract features from the input sequence and extract the multimodal features of the signal in different ways; the feature fusion module is used to fuse the multimodal features into more informative fused features through the attention mechanism; the classifier is used to further process the fused features to obtain the modulation mode of the final perceived signal.

[0034] The beneficial effects of the present invention include:

[0035] The radio signal waveform perception method in the present invention makes full use of the time domain, frequency domain and other characteristics of the radio signal, forms multi-dimensional features by mapping to the real domain and complex domain, and realizes the refined extraction of multi-domain features. The attention mechanism is designed to deeply fuse the multi-domain characteristics, and the accuracy of target signal perception is improved by establishing an accurate matching relationship between the fused features and the radio signal waveform characteristics (such as the modulation mode). BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0037] Figure 1 It is a schematic diagram of the overall architecture of an embodiment of the present invention;

[0038] Figure 2 A schematic diagram of the structure of a feature fusion module according to an embodiment of the present invention;

[0039] Figure 3 Schematic diagram of the attention mechanism structure of an embodiment of the present invention;

[0040] Figure 4 Schematic diagram of the residual module structure according to an embodiment of the present invention;

[0041] Figure 5 Schematic diagram of the residual unit structure of an embodiment of the present invention;

[0042] Figure 6 Schematic diagram for ablation experiment comparison;

[0043] Figure 7 A schematic diagram comparing an embodiment of the present invention with an existing modulation sensing algorithm;

[0044] Figure 8 Schematic diagram of the effect of different numbers of sampling points on the accuracy of an embodiment of the present invention. DETAILED DESCRIPTION

[0045] All features disclosed in all embodiments in this specification, or steps in all methods or processes implicitly disclosed, except for mutually exclusive features and / or steps, can be combined and / or expanded or replaced in any manner.

[0046] In view of the problems in the background, the present invention proposes a fusion multi-domain radio signal waveform perception method based on the prior art, aiming to fully extract the time domain and frequency domain features of the radio signal, map to form real and complex multi-mode features, and efficiently fuse the multi-mode features, thereby improving the accuracy of radio signal waveform perception. The specific implementation process of the present invention is as follows:

[0047] Step A Input feature description

[0048] The present invention is mainly aimed at detecting non-cooperative transmitters, which are related to the transmission of communication signals and radar signals. The present invention takes the time domain sequence and frequency domain sequence of signal sampling as input. The time domain sequence of the received signal is expressed as:

[0049] r(t)=x(t)*h(t)+n(t),

[0050] Among them, x(t) represents the modulated signal transmitted by the dual-function transmitter, h(t) represents the impulse response of the channel, and n(t) represents the additive Gaussian noise. After receiving the signal, the receiver converts the signal into a discrete form r[n], which is usually expressed as two sets of orthogonal components, namely

[0051] r[n]=r I [n]+jr Q [n],

[0052] Where n represents the sampling number, j represents the imaginary unit and j 2 =-1, r I [n] represents the in-phase (I) component, r Q [n] represents the quadrature (Q) component.

[0053] The frequency domain sequence of a signal is usually obtained by discrete Fourier transform (DFT), which can be expressed as:

[0054]

[0055] Where N represents the length of DFT, There is a fast implementation method for the DFT algorithm, namely Fast Fourier Transform (FFT). FFT can effectively reduce the computational complexity of frequency domain transformation.

[0056] The present invention uses time domain IQ sequence and FFT sequence as input features, which has the advantages that the IQ sequence is a native signal directly received by the receiver and does not require additional processing; the FFT algorithm currently has a mature and complete strategy and is relatively simple to implement.

[0057] Step B: Design of intelligent modulation sensing technology solution

[0058] As mentioned above, the present invention uses IQ and FFT sequences as input. Figure 1 As shown, the present invention is mainly divided into three parts: feature extraction module, feature fusion module and classifier. Specifically, the feature extraction module extracts features from the input sequence and extracts the multi-modal features of the signal in different ways; the feature fusion module fuses the multi-modal features into more informative fusion features through the attention mechanism; the classifier further processes the fusion features and finally perceives the modulation mode of the signal.

[0059] S1. Feature extraction module structure: The feature extraction module performs real value convolution (RVC) and complex value convolution (CVC) operations on the two sets of sequences respectively. Since the processing methods of the two sets of data are the same, the present invention only describes the processing method of the time domain sequence T in detail.

[0060] In neural networks, all calculations are based on real numbers, and common convolutions belong to the category of RVC. The sequence T is a complex vector, and the CVC operation is defined as

[0061] T*h=(T r +jT i )*(h r +jh i )=(T r *h r -T i *h i )+j(T r *h i +T i *h r ),

[0062] Among them, T rand T i represent the real and imaginary parts of the time domain sequence T, h is the complex convolution kernel, and h r and h c Represent the real and imaginary parts of h respectively. In neural networks, complex numbers are usually represented by real and imaginary parts, so the CVC operation is equivalent to

[0063]

[0064] It can be seen that the CVC operation requires two real convolution kernels h r and h i , then the CVC operation can be obtained by two RVC operations.

[0065] The RVC operation focuses on extracting individual features of the real or imaginary part of the input sequence. Theoretically, it is impossible to extract the correlation information between the real and imaginary parts. CVC makes up for this shortcoming.

[0066] By performing RVC and CVC operations on the time domain sequence T and the frequency domain sequence F respectively, four sets of feature flows can be obtained. Figure 1 In the figure, they are represented by T R , T C , F R and F C .

[0067] S2. Feature fusion module structure: The four sets of feature streams extracted by the feature extraction module represent different features in the time domain and frequency domain. In order to improve the performance of modulation perception, it is necessary to fuse the features of multiple feature streams instead of simply splicing them according to a certain dimension. Figure 2 As shown in Figure 1, the feature fusion module mainly consists of a depthwise separable convolution (DSC) layer and an attention module.

[0068] The DSC layer consists of channel-dimensional convolution and 1×1 convolution. Compared with the traditional convolution layer, channel-dimensional convolution can effectively reduce the number of parameters for each convolution kernel and effectively reduce calculations. The transformation of the output feature map dimension is completed through 1×1 convolution. The DSC layer splits the convolution layer into two parts, which can effectively reduce the number of parameters to be trained.

[0069] The attention module uses the channel attention mechanism to automatically correct the feature strength between channels by modeling the interdependence between channels. Its structure is as follows Figure 3As shown. First, the feature map is squeezed and the global average is calculated according to the channel dimension. Then the weights corresponding to each channel of the feature map are obtained through the activation function Sigmoid, and finally the feature map is multiplied by the corresponding weight to obtain the fused feature. The channel attention mechanism corrects the strength of the feature by assigning corresponding weights to different channels of the feature map. Compared with the input features, the important information in the fused features is more prominent and has stronger representation ability. It should be noted that in order to prevent the gradient from disappearing, there is a jump connection between the input and output of the attention mechanism.

[0070] S3, classifier structure: After receiving the fused features, the residual module and the fully connected layer module are used to further process the fused features and map them to the perception results. Figure 4 As shown, the residual module consists of two residual units and a pooling layer. The residual unit structure is as follows Figure 5 As shown in Figure 2. Each residual unit consists of two RVC layers and a skip connection. The skip connection can be expressed as

[0071] FF′=FF+f(FF),

[0072] Among them, FF represents the input feature map, f(FF) represents the feature map after two RVC operations, and FF′ represents the output feature map after the jump connection. During back propagation, even if the gradient of the convolution output f(FF) to the input feature map FF is 0, the jump connection output FF′ can ensure that the gradient to FF is at least 1, thereby avoiding the gradient vanishing phenomenon. The role of the pooling layer is to reduce the number of output neurons to prevent overfitting.

[0073] The fully connected layer module first flattens the feature map into a one-dimensional vector, and then inputs the vector into the fully connected layer, i.e.

[0074] f k =f act (w k f k-1 )+b k ,k=1,2,…n,

[0075] Where n represents the number of fully connected layers, f k represents the output of the kth fully connected layer, w k and b k are the weight and bias of the kth fully connected layer respectively.

[0076] Step C: Training Intelligent Modulation Perception Technology

[0077] SS1, Dataset:

[0078] The communication signal modulation modes considered in the present invention include BPSK (Binary Phase Shift Keying), QPSK (Quadrature PSK), 8PSK, CPFSK (Continuous Phase Frequency Shift Keying), GFSK (Gaussian FSK), QAM16 (Quadrature Amplitude Modulation16), QAM64 and AM (Analog modulation), and the radar signal modulation modes include Linear Frequency Modulation (LFM) and Pulse Compression Modulation (PCM). The signal-to-noise ratio ranges from -5dB to 15dB. Considering the relative motion between the transmitter and the receiver, the maximum Doppler shift is set to 20kHz.

[0079] SS2, loss function selection:

[0080] The present invention uses the SS1 data set in C to train the intelligent modulation perception network constructed in step B, and uses the cross entropy loss as the loss function to obtain the training loss, that is,

[0081]

[0082] Where y represents the actual modulation mode of the signal, represents the modulation mode perceived by the present invention. When the actual modulation mode is the i-th type, y i is 1, otherwise it is 0. By continuously minimizing the loss function, the predicted value can be gradually approached to the true value, thereby achieving training.

[0083] SS3, Perceptual Modulation Selection:

[0084] The receiver inputs the received signal into the network constructed in B to obtain the calculation result of modulation perception. The index with the highest calculation probability result is used as the modulation mode corresponding to the received signal, that is,

[0085]

[0086] The following is a combination of simulation examples and drawings to illustrate the technical effects of the present invention. First, the data set parameters and system model parameter settings are provided, and then the simulation results are provided to evaluate the performance of the present invention. At the same time, the proposed algorithm is compared with the existing modulation perception algorithm, and the influence of different sampling points on the perception accuracy of the present invention is analyzed.

[0087] The present invention considers 8 communication signals, including BPSK, QPSK, 8PSK, CPFSK, GFSK, QAM16, QAM64 and AM, and 2 radar signals, LFM and PCM. The baseband bandwidth is 100MHz, the carrier frequency is 2.4GHz, and the upsampling multiple is 8. The channel environment is a Gaussian channel, and the simulated signal-to-noise ratio range is -5dB to 15dB. Considering the relative motion between the transmitter and the receiver, the simulated maximum Doppler frequency deviation is 20KHz.

[0088] The neural network model parameters of the present invention are set as follows. Figure 1 The overall structure of the algorithm shown in the figure is that a complex convolution consists of two real convolution layers, where the size of the real convolution kernel is (1, 3), Relu is used as the activation function, and the output channels are 16. In the DSC layer in the feature fusion module, the convolution kernel size of the channel dimension convolution is (1, 3), and the number of output channels of the convolution layer is 16. In the residual module, the convolution kernel size of the real convolution is set to (1, 3), and the number of output channels is 16. In the above convolution operations, except for convolution, the rest use zero padding operations to keep the dimension of the feature map constant. The size of the pooling layer is set according to the actual situation, and the step size is (2, 2). After the flattening layer, the feature vector obtained by the above convolution layer is converted into a one-dimensional vector. The length of the hidden layer in the fully connected layer is 2, the number of neurons is 128 and 64 respectively, Selu is used as the activation function, the number of neurons in the output layer is 10, and the activation function is Softmax. The network uses the cross entropy loss function, the optimizer is Adam, and the initial learning rate is 0.001. At each signal-to-noise ratio, the present invention collects 2000 frames of each modulated signal, and each frame signal has 128 sampling points. The ratio of the training set, the validation set, and the test set is set according to the actual situation.

[0089] Figure 6The comparison results of the ablation experiment are shown. There are 5 comparison groups in the ablation experiment, namely, the removal of time domain information, the removal of frequency domain information, the removal of complex convolution, the removal of real convolution and the removal of attention module groups of the present invention. In order to ensure the fairness of the experiment, the number of convolution kernels in the ablation experiment group is doubled to ensure that the number of parameters in the feature processing stage is equivalent and the dimension of the feature map output by the splicing layer is consistent. The present invention observes that the accuracy of the frequency domain information removal group converges to about 94%, while the time domain information removal group only converges to 85%. It can be inferred that for the modulation perception task, time domain information is more important than frequency domain information, and the rational use of the two types of information can further improve the accuracy. Compared with the removal of the frequency domain information group, the perception accuracy of the present invention is improved by about 6%. The perception accuracy of the real convolution group removed converges to about 92%, and the perception accuracy of the complex convolution group removed converges to 87%. It can be inferred that the complex features of the signal contain more information than the individual features of the real and imaginary parts, and the fusion of the two features can further improve the perception accuracy. In addition, the perception accuracy of the group without the attention mechanism is about 2% lower than that of the present invention, which shows that the attention mechanism can help the neural network focus on important information and further improve the perception accuracy. Figure 6 This shows the effectiveness of the multi-mode features and feature extraction method used in the present invention.

[0090] Figure 7 The schematic diagram of the comparison between the present invention and the existing modulation sensing technology algorithm is shown. In the range of -5dB to 3dB, the performance of the present invention is comparable to that of the existing modulation sensing technology. In the range of 4dB to 15dB, the modulation sensing accuracy of the present invention is about 4% higher than that of the existing technology on average. Figure 7 This shows that the technology proposed in the present invention has an advantage in perception accuracy.

[0091] Figure 8 The influence of different numbers of sampling points on the accuracy of the present invention is demonstrated. As the number of sampling points increases, the modulation perception accuracy also increases. Among them, under the same signal-to-noise ratio, the recognition accuracy of 128 points, 256 points, and 512 points is within 5%. At -5dB to 5dB, the recognition accuracy of 64 points is 10% lower than that of 128 points on average. Taking all factors into consideration, the present invention uses 128 sampling points as input.

[0092] It should be noted that within the scope of protection defined in the claims of the present invention, the following embodiments can be combined and / or expanded or replaced in any logical way from the above specific implementation methods, such as disclosed technical principles, disclosed technical features or implicitly disclosed technical features.

[0093] Example 1

[0094] A radio signal waveform intelligent perception method integrating multi-domain features, comprising:

[0095] Perform feature extraction on the input sequence and extract the multi-modal features of the signal in different ways; then fuse the multi-modal features into more informative fused features through the attention mechanism;

[0096] Finally, the fused features are further classified and processed to obtain the modulation mode of the final perception signal.

[0097] Example 2

[0098] Based on Example 1, the input sequence includes IQ and FFT sequences.

[0099] Example 3

[0100] On the basis of Example 1, the feature extraction of the input sequence extracts the multi-mode features of the signal in different ways, specifically including the sub-steps of performing real convolution RVC and complex convolution CVC operations on two groups of sequences respectively, so as to obtain multiple groups of feature streams.

[0101] Example 4

[0102] On the basis of Example 1, the method of fusing multimodal features into more informative fusion features through the attention mechanism specifically includes the following sub-steps:

[0103] The convolution layer is split into two parts using the depthwise separable convolutional layer DSC, which consists of channel dimension convolution and 1×1 convolution. The transformation of the output feature map dimension is completed by 1×1 convolution.

[0104] In addition, the channel attention mechanism is used to correct the feature strength between channels by modeling the interdependence between channels. Specifically, the feature map is first compressed and the global average is calculated according to the channel dimension; then the weight corresponding to each channel of the feature map is obtained through the activation function, and finally the feature map is multiplied by the corresponding weight to obtain the fused feature.

[0105] Example 5

[0106] Based on Example 4, the activation function is a Sigmoid activation function.

[0107] Example 6

[0108] Based on Example 4, the method further includes a sub-step of providing a jump connection between the input and output of the attention mechanism to prevent the gradient from disappearing.

[0109] Example 7

[0110] On the basis of Example 1, the fusion features are further classified and processed to obtain the modulation method of the final perception signal, which specifically includes sub-steps: after receiving the fusion features, the fusion features are further processed using a residual module and a fully connected layer module, and mapped to the perception results.

[0111] Example 8

[0112] On the basis of Example 7, after receiving the fused features, the fused features are further processed using a residual module and a fully connected layer module, and mapped to the perception results, specifically including the following sub-steps:

[0113] The residual module consists of two residual units and a pooling layer. Each residual unit consists of two RVC layers and a skip connection. The skip connection is represented as:

[0114] FF′=FF+f(FF),

[0115] Among them, FF represents the input feature map, f(FF) represents the feature map after two RVC operations, and FF′ represents the output feature map after the jump connection. During back propagation, even if the gradient of the convolution output f(FF) to the input feature map FF is 0, the jump connection output FF′ can ensure that the gradient to FF is at least 1, which is used to avoid the gradient disappearance phenomenon. The pooling layer is used to reduce the number of output neurons to prevent overfitting.

[0116] The fully connected layer module first flattens the feature map into a one-dimensional vector, and then inputs the vector into the fully connected layer, namely:

[0117] f k =f act (w k f k-1 )+b k ,k=1,2,…n,

[0118] Among them, n represents the number of fully connected layers, f k represents the output of the kth fully connected layer, w k and b k are the weight and bias of the kth fully connected layer respectively.

[0119] Example 9

[0120] On the basis of Example 1, the feature extraction of the input sequence extracts the multi-mode features of the signal in different ways, specifically including: extracting the time domain and frequency domain features of the radio signal, and mapping them to form real and complex multi-mode features.

[0121] Example 10

[0122] A radio signal waveform intelligent perception system integrating multi-domain features, comprising a feature extraction module, a feature fusion module and a classifier;

[0123] The feature extraction module is used to extract features from the input sequence and extract the multimodal features of the signal in different ways; the feature fusion module is used to fuse the multimodal features into more informative fused features through the attention mechanism; the classifier is used to further process the fused features to obtain the modulation mode of the final perceived signal.

[0124] The units involved in the embodiments of the present invention may be implemented by software or hardware, and the units described may also be arranged in a processor. The names of these units do not, in some cases, limit the units themselves.

[0125] According to one aspect of an embodiment of the present invention, a computer program product or a computer program is provided, the computer program product or the computer program includes a computer instruction, and the computer instruction is stored in a computer-readable storage medium. A processor of a computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the method provided in the above various optional implementations.

[0126] As another aspect, an embodiment of the present invention further provides a computer-readable medium, which may be included in the electronic device described in the above embodiment; or may exist independently without being assembled into the electronic device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by an electronic device, the electronic device implements the method described in the above embodiment.

Claims

1. A radio signal waveform intelligent perception method integrating multi-domain features, characterized in that: include: Perform feature extraction on the input sequence and extract the multi-mode features of the signal in different ways; Then, the multimodal features are fused into more informative fused features through the attention mechanism; Finally, the fused features are further classified and processed to obtain the modulation mode of the final perception signal.

2. The method for intelligently sensing radio signal waveforms by integrating multi-domain features according to claim 1 is characterized in that: The input sequence includes IQ and FFT sequences.

3. The method for intelligently sensing radio signal waveforms by integrating multi-domain features according to claim 1 is characterized in that: The feature extraction of the input sequence extracts the multi-mode features of the signal in different ways, and specifically includes the sub-steps of performing real convolution RVC and complex convolution CVC operations on two groups of sequences respectively, so as to obtain multiple groups of feature streams.

4. The method for intelligently sensing radio signal waveforms by integrating multi-domain features according to claim 1 is characterized in that: The method of fusing multi-modal features into more informative fusion features through the attention mechanism specifically includes the following sub-steps: The convolution layer is split into two parts using the depthwise separable convolutional layer DSC, which consists of channel dimension convolution and 1×1 convolution. The transformation of the output feature map dimension is completed by 1×1 convolution. In addition, the channel attention mechanism is used to correct the feature strength between channels by modeling the interdependence between channels. Specifically, the feature map is first compressed and the global average is calculated according to the channel dimension; then the weight corresponding to each channel of the feature map is obtained through the activation function, and finally the feature map is multiplied by the corresponding weight to obtain the fused feature.

5. The method for intelligently sensing radio signal waveforms by integrating multi-domain features according to claim 4 is characterized in that: The activation function is a Sigmoid activation function.

6. The method for intelligently sensing radio signal waveforms by integrating multi-domain features according to claim 4 is characterized in that: The method also includes a sub-step of providing a jump connection between the input and output of the attention mechanism to prevent the gradient from disappearing.

7. The method for intelligently sensing radio signal waveforms by integrating multi-domain features according to claim 1 is characterized in that: The further classification processing of the fused features to obtain the modulation mode of the final perception signal specifically includes the following sub-steps: after receiving the fused features, further processing the fused features using a residual module and a fully connected layer module, and mapping them to the perception results.

8. The method for intelligently sensing radio signal waveforms by integrating multi-domain features according to claim 7 is characterized in that: After receiving the fused features, the fused features are further processed using a residual module and a fully connected layer module, and mapped to the perception results, specifically including the following sub-steps: The residual module consists of two residual units and a pooling layer. Each residual unit consists of two RVC layers and a skip connection. The skip connection is represented as: FF′=FF+f(FF), Among them, FF represents the input feature map, f(FF) represents the feature map after two RVC operations, and FF′ represents the output feature map after the jump connection. During back propagation, even if the gradient of the convolution output f(FF) to the input feature map FF is 0, the jump connection output FF′ can ensure that the gradient to FF is at least 1, which is used to avoid the gradient disappearance phenomenon. The pooling layer is used to reduce the number of output neurons to prevent overfitting. The fully connected layer module first flattens the feature map into a one-dimensional vector, and then inputs the vector into the fully connected layer, namely: f k =f act (w k f k-1 )+b k ,k=1,2,…n, Among them, n represents the number of fully connected layers, f k represents the output of the kth fully connected layer, w k and b k are the weight and bias of the kth fully connected layer respectively.

9. The method for intelligently sensing radio signal waveforms by integrating multi-domain features according to claim 1 is characterized in that: The feature extraction of the input sequence extracts the multi-mode features of the signal in different ways, specifically including: extracting the time domain and frequency domain features of the radio signal, and mapping them to form real and complex multi-mode features.

10. A radio signal waveform intelligent perception system integrating multi-domain features, characterized in that: Includes feature extraction module, feature fusion module and classifier; The feature extraction module is used to extract features from the input sequence and extract the multimodal features of the signal in different ways; the feature fusion module is used to fuse the multimodal features into more informative fused features through the attention mechanism; the classifier is used to further process the fused features to obtain the modulation mode of the final perceived signal.

Images