A radio signal waveform intelligent perception method and system fusing multi-domain features

Abstract

The application discloses a wireless telecommunication signal waveform intelligent sensing method and system fusing multi-domain features, and belongs to the field of wireless communication, and comprises the following steps: feature extraction is performed on an input sequence, and multi-mode features of a signal are extracted through different modes; then, the multi-mode features are fused into fused features with richer information through an attention mechanism; and finally, the fused features are further classified to obtain a modulation mode of a final sensing signal. The application can improve the accuracy of target signal sensing.

Description

Technical Field

[0001] This invention relates to the field of wireless communication, and more specifically, to a method and system for intelligent sensing of radio signal waveforms that integrates multi-domain features. Background Technology

[0002] Targets typically carry multiple devices such as communication and radar systems to ensure mission execution. Depending on the mission, targets will emit different radio signal waveforms; for example, communication and radar terminals may select different frequencies, bandwidths, and modulation methods. By sensing the target's radio signals, the target's operational status can be analyzed, providing support for determining the target's intentions.

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

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

[0005] (2) Feature-based methods. This method extracts various 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, feature selection relies heavily on manual intervention, and its performance is poor when identifying high-order modulated radio signals.

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

[0007] (1) Architecture based on Convolutional Neural Network (CNN). CNN extracts modulation pattern features from radio signal waveforms by performing convolution operations on the sequence of received signals or transform domain images. As the number of convolutional layers increases, the receptive field of the output neurons also increases, and the representational power of the output feature map is usually stronger.

[0008] (2) Architecture based on Complex Value Convolution Neural Network (CVCNN). The CVCNN architecture typically processes complex sequences of signals. Similar to 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) Architecture based on Long Short-Term Memory (LSTM). LSTM architecture typically processes time-domain sequences of signals, sensing the modulation mode of the signal through the temporal 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 these methods is used, the radio signal waveform sensing performance is often poor. Method (3) cannot achieve complete parallelism in structure due to the use of the time correlation of the input, resulting in high latency. Summary of the Invention

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

[0012] The objective of this invention is achieved through the following solution:

[0013] A method for intelligent sensing of radio signal waveforms that integrates multi-domain features, comprising:

[0014] Feature extraction is performed on the input sequence, extracting multi-modal features of the signal through different methods; then, an attention mechanism is used to fuse the multi-modal features into a more informative fused feature.

[0015] Finally, the fused features are further classified to obtain the modulation scheme of the final perceived signal.

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

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

[0018] Furthermore, the process of fusing multi-modal features into richer fused features through an attention mechanism specifically includes the following sub-steps:

[0019] The depthwise separable convolutional layer (DSC) is used to split the convolutional layer into two parts, specifically consisting of channel-dimensional convolution and 1×1 convolution. The transformation of the output feature map dimension is accomplished through 1×1 convolution.

[0020] Furthermore, a channel attention mechanism is adopted to correct the feature strength between channels by modeling the interdependencies between channels. Specifically, this includes: firstly, compressing the feature map and calculating the global average according to the channel dimension; then obtaining the weights corresponding to each channel of the feature map through the activation function; and finally multiplying the feature map with the corresponding weights to obtain the fused features.

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

[0022] Furthermore, it also includes a sub-step: a skip connection is provided between the input and output of the attention mechanism to prevent gradient vanishing.

[0023] Furthermore, the step of further classifying the fused features to obtain the modulation scheme of the final sensing signal specifically includes the following sub-steps: 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 sensing result.

[0024] Furthermore, 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 result, specifically including the following sub-steps:

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

[0026] FF′=FF+f(FF),

[0027] Here, 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 skip connection. During backpropagation, even if the gradient of the convolution output f(FF) with respect to the input feature map FF is 0, the gradient of the skip connection output FF′ with respect to FF is guaranteed to be at least 1, which is used to avoid the phenomenon of gradient vanishing. The pooling layer is used to reduce the number of output neurons and prevent overfitting.

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

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

[0030] Where n represents the number of fully connected layers, f k w represents the output of the k-th fully connected layer. k and b kThese represent the weights and biases of the k-th fully connected layer, respectively.

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

[0032] A radio signal waveform intelligent sensing system that integrates multi-domain features includes 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 multi-modal features of the signal in different ways; the feature fusion module is used to fuse multi-modal features into more informative fused features through an attention mechanism; the classifier further processes the fused features to obtain the modulation scheme of the final perceived signal.

[0034] The beneficial effects of this invention include:

[0035] The radio signal waveform sensing method in this invention fully utilizes the time and frequency domain features of radio signals, mapping them to the real and complex domains to form multi-dimensional features, thereby achieving refined extraction of multi-domain features. An attention mechanism is designed to deeply fuse multi-domain characteristics, improving the accuracy of target signal sensing by establishing a precise matching relationship between the fused features and the waveform characteristics of the radio signal (such as modulation scheme). Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

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

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

[0039] Figure 3 This is a schematic diagram of the attention mechanism structure according to an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram of the residual module structure according to an embodiment of the present invention;

[0041] Figure 5 This is a schematic diagram of the residual unit structure according to an embodiment of the present invention;

[0042] Figure 6 This is a schematic diagram comparing ablation experiments;

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

[0044] Figure 8 This diagram illustrates the impact of different numbers of sampling points on the accuracy of embodiments of the present invention. Detailed Implementation

[0045] All features disclosed in all embodiments of this specification, or steps in all methods or processes implied in the disclosure, may be combined and / or extended or replaced in any way, except for mutually exclusive features and / or steps.

[0046] In view of the problems mentioned in the background, this invention proposes a multi-domain radio signal waveform sensing method based on existing technologies. The method aims to fully extract the time-domain and frequency-domain features of radio signals, map them to form real-number and complex-number multi-mode features, and efficiently fuse these multi-mode features, thereby improving the accuracy of radio signal waveform sensing. The specific implementation process of this invention is as follows:

[0047] Step A: Input Feature Description

[0048] This invention primarily targets the detection of non-cooperative transmitters involved in transmitting communication and radar signals. The invention uses the time-domain and frequency-domain sequences of the sampled signal as input. The time-domain sequence of the received signal is represented as follows:

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

[0050] Where 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 additive Gaussian noise. After receiving the signal, the receiver converts it into a discrete form r[n], which is usually represented by two sets of orthogonal components, i.e.

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

[0052] Where n represents the sampling sequence 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 the Discrete Fourier Transform (DFT), and can be represented as:

[0054]

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

[0056] This invention uses time-domain IQ sequences and FFT sequences as input features. Its advantages are: the IQ sequence is the native signal directly received by the receiver and does not require additional processing; the FFT algorithm currently has mature and well-established strategies and is relatively simple to implement.

[0057] Step B: Design of Intelligent Modulation Sensing Technology Solution

[0058] As described above, this invention uses IQ and FFT sequences as input. Figure 1 As shown, this invention mainly consists of three parts: a feature extraction module, a feature fusion module, and a classifier. Specifically, the feature extraction module extracts features from the input sequence by extracting multi-modal features of the signal through different methods; the feature fusion module fuses the multi-modal features into a more informative fused feature through an attention mechanism; and the classifier further processes the fused feature to ultimately perceive 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 for the two sets of data are the same, this invention only describes in detail the processing method for the time-domain sequence T.

[0060] In neural networks, all computations are based on real numbers, and common convolutions fall under the category of Recurrent Vector Convolution (RVC). A sequence T is a complex vector, and the Concurrent Vector Convolution (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 Representing the real and imaginary parts of the time-domain sequence T, respectively, and h is the complex convolution kernel. r and h c Let represent the real and imaginary parts of h, respectively. In neural networks, complex numbers are typically represented by real and imaginary parts, so the CVC operation is equivalent to...

[0063]

[0064] Therefore, CVC operation requires two real convolution kernels h. r and h i Then the CVC operation can be obtained by two RVC operations.

[0065] RVC focuses on extracting individual features of the real or imaginary part of the input sequence, and theoretically cannot extract the correlation information between the real and imaginary parts. CVC, on the other hand, 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 T is represented as T. R T C F R and F C .

[0067] S2. Feature Fusion Module Structure: The four feature streams extracted by the feature extraction module represent different features in the time and frequency domains. To improve the performance of modulation sensing, feature fusion of multiple feature streams is necessary, rather than simply concatenating them according to a certain dimension. For example... Figure 2 As shown, the feature fusion module mainly consists of a depthwise separable convolution (DSC) layer and an attention module.

[0068] The DSC layer consists of channel-wise convolutions and 1×1 convolutions. Compared to traditional convolutional layers, channel-wise convolutions effectively reduce the number of parameters in each convolutional kernel and thus reduce computation. The transformation of the output feature map dimensions is accomplished through 1×1 convolutions. By splitting the convolutional layer into two parts, the DSC layer effectively reduces the number of parameters to be trained.

[0069] The attention module employs a channel attention mechanism, which automatically corrects the feature intensity between channels by modeling the interdependencies between them. Its structure is as follows: Figure 3As shown, the feature map is first compressed (squeeze) and a global average is calculated along the channel dimension. Then, the weights corresponding to each channel of the feature map are obtained through the sigmoid activation function. Finally, the feature map is multiplied by its corresponding weights to obtain the fused feature. The channel attention mechanism corrects the strength of the features by assigning appropriate weights to different channels of the feature map. Compared with the input features, the important information in the fused feature is more prominent, and the representational ability is stronger. It is important to note that to prevent gradient vanishing, there are skip connections between the input and output of the attention mechanism.

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

[0071] FF′=FF+f(FF),

[0072] Here, 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 skip connection. During backpropagation, even if the gradient of the convolutional output f(FF) with respect to the input feature map FF is 0, the skip connection output FF′ ensures that its gradient with respect to FF is at least 1, thus avoiding the vanishing gradient phenomenon. The pooling layer reduces 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 this vector into the fully connected layer.

[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 w represents the output of the k-th fully connected layer. k and b k These represent the weights and biases of the k-th fully connected layer, respectively.

[0076] Step C: Training Intelligent Modulation Sensing Technology

[0077] SS1, Dataset:

[0078] The communication signal modulation methods considered in this invention include BPSK (Binary Phase Shift Keying), QPSK (Quadrature PSK), 8PSK, CPFSK (Continuous Phase Frequency Shift Keying), GFSK (Gaussian FSK), QAM16 (Quadrature Amplitude Modulation 16), QAM64, and AM (Analog modulation). Radar signal modulation methods include Linear Frequency Modulation (LFM) and Pulse Compression Modulation (PCM). The signal-to-noise ratio range is -5dB to 15dB. Considering the relative motion between the transmitter and receiver, the maximum Doppler frequency shift is set to 20kHz.

[0079] SS2, Loss Function Selection:

[0080] This invention uses the SS1 dataset from C to train the intelligent modulation sensing network constructed in step B, and uses cross-entropy loss as the loss function to obtain the training loss, i.e.

[0081]

[0082] Where y represents the actual modulation scheme of the signal. This represents the modulation scheme perceived by the present invention. When the actual modulation scheme is the i-th type, y i The value is 1 if the error condition is true and 0 otherwise. By continuously minimizing the loss function, the predicted value can be gradually made closer to the true value, thus achieving training.

[0083] SS3, Sensing Modulation Method Selection:

[0084] The receiver inputs the received signal into the network constructed in B to obtain the modulation sensing calculation results. The index with the highest calculated probability result is taken as the modulation scheme corresponding to the received signal, i.e.

[0085]

[0086] The technical effects of this invention are illustrated below with simulation examples and accompanying figures. First, the dataset parameters and system model parameter settings are provided. Then, simulation results are presented to evaluate the performance of this invention. Furthermore, the proposed algorithm is compared with existing modulation sensing algorithms, and the impact of different sampling point numbers on the sensing accuracy of this invention is analyzed.

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

[0088] The neural network model parameters of this invention are set as follows. For example... Figure 1 The algorithm's overall structure is shown. A complex convolution consists of two real convolutional layers, where the kernel size is (1, 3), ReLU is used as the activation function, and both layers have 16 output channels. In the DSC layer of the feature fusion module, the kernel size for the channel-dimensional convolution is (1, 3), and the number of output channels is 16. In the residual module, the kernel size for the real convolution is set to (1, 3), and the number of output channels is 16. In all the above convolutional operations, except for convolution, zero-padding is used to maintain the constant dimension of the feature map. The pooling layer size is set according to the actual situation, with a stride of (2, 2). After the flattening layer, the feature vectors obtained from the above convolutional layers are transformed into one-dimensional vectors. The hidden layer length in the fully connected layer is 2, with 128 and 64 neurons respectively, using SeLU as the activation function. The output layer has 10 neurons, and the activation function is Softmax. The network uses cross-entropy as the loss function, Adam as the optimizer, and an initial learning rate of 0.001. At each signal-to-noise ratio, this invention collects 2000 frames of each modulated signal, with 128 sampling points per frame. The ratio of training set, validation set, and test set is set according to the actual situation.

[0089] Figure 6The comparison results of the ablation experiments are presented. Five comparison groups were conducted: those for removing temporal information, removing frequency domain information, removing complex convolutions, removing real convolutions, and removing the attention module. To ensure fairness, the number of convolutional kernels in the ablation experiment groups was doubled to ensure a comparable number of parameters in the feature processing stage and to ensure consistent dimensionality of the feature maps output by the splicing layer. The invention observed that the accuracy of the group removing frequency domain information converged to approximately 94%, while the group removing temporal information converged only to 85%. This suggests that for modulation sensing tasks, temporal information is more important than frequency domain information, and that the reasonable utilization of both types of information can further improve accuracy. Compared to the group removing frequency domain information, the invention improved sensing accuracy by approximately 6%. The sensing accuracy of the group removing real convolutions converged to approximately 92%, while the sensing accuracy of the group removing complex convolutions converged to 87%. This suggests that the complex features of a signal contain richer information than the individual features of the real and imaginary parts, and that fusing both types of features can further improve sensing accuracy. Furthermore, the perceptual accuracy of the group without the attention mechanism was about 2% lower than that of the present invention, indicating that the attention mechanism can help the neural network focus on important information and further improve the perceptual accuracy. Figure 6 This demonstrates the effectiveness of the multi-modal feature and feature extraction methods employed in this invention.

[0090] Figure 7 A schematic diagram comparing the present invention with existing modulation sensing algorithms is shown. Within the range of -5dB to 3dB, the performance of the present invention is comparable to existing modulation sensing technologies. Within the range of 4dB to 15dB, the modulation sensing accuracy of the present invention is approximately 4% higher on average than that of existing technologies. Figure 7 This demonstrates the advantage of the technology proposed in this invention in terms of perception accuracy.

[0091] Figure 8 The effect of different sampling point numbers on the accuracy of this invention is demonstrated. As the number of sampling points increases, the modulation sensing accuracy also increases. Specifically, under the same signal-to-noise ratio, the accuracy differences for 128 points, 256 points, and 512 points are all within 5%. Within the -5dB to 5dB range, the accuracy of 64 points is on average 10% lower than that of 128 points. Taking all factors into consideration, this invention uses 128 sampling points as the input.

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

[0093] Example 1

[0094] A method for intelligent sensing of radio signal waveforms that integrates multi-domain features, comprising:

[0095] Feature extraction is performed on the input sequence, extracting multi-modal features of the signal through different methods; then, an attention mechanism is used to fuse the multi-modal features into a more informative fused feature.

[0096] Finally, the fused features are further classified to obtain the modulation scheme of the final perceived signal.

[0097] Example 2

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

[0099] Example 3

[0100] Based on Example 1, the feature extraction of the input sequence is carried out by extracting multi-mode features of the signal in different ways. Specifically, it includes the following sub-steps: performing real convolution (RVC) and complex convolution (CVC) operations on the two sets of sequences respectively to obtain multiple sets of feature streams.

[0101] Example 4

[0102] Based on Example 1, the process of fusing multi-modal features into richer fused features through an attention mechanism specifically includes the following sub-steps:

[0103] The depthwise separable convolutional layer (DSC) is used to split the convolutional layer into two parts, specifically consisting of channel-dimensional convolution and 1×1 convolution. The transformation of the output feature map dimension is accomplished through 1×1 convolution.

[0104] Furthermore, a channel attention mechanism is adopted to correct the feature strength between channels by modeling the interdependencies between channels. Specifically, this includes: firstly, compressing the feature map and calculating the global average according to the channel dimension; then obtaining the weights corresponding to each channel of the feature map through the activation function; and finally multiplying the feature map with the corresponding weights to obtain the fused features.

[0105] Example 5

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

[0107] Example 6

[0108] Based on Example 4, a sub-step is also included: a skip connection is provided between the input and output of the attention mechanism to prevent gradient vanishing.

[0109] Example 7

[0110] Based on Example 1, the further classification and processing of the fused features to obtain the modulation method of the final sensing signal specifically includes the following sub-steps: 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 sensing result.

[0111] Example 8

[0112] Based on Example 7, the step of further processing the fused features using a residual module and a fully connected layer module after receiving the fused features, and mapping them to the perception result, specifically includes the following sub-steps:

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

[0114] FF′=FF+f(FF),

[0115] Here, 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 skip connection. During backpropagation, even if the gradient of the convolution output f(FF) with respect to the input feature map FF is 0, the gradient of the skip connection output FF′ with respect to FF is guaranteed to be at least 1, which is used to avoid the phenomenon of gradient vanishing. The pooling layer is used to reduce the number of output neurons and prevent overfitting.

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

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

[0118] Where n represents the number of fully connected layers, f k w represents the output of the k-th fully connected layer. k and b k These represent the weights and biases of the k-th fully connected layer, respectively.

[0119] Example 9

[0120] Based on Example 1, the feature extraction of the input sequence is carried out in different ways to extract the multi-mode features of the signal, specifically including: extracting the time domain and frequency domain features of the radio signal and mapping them to form real number and complex number multi-mode features.

[0121] Example 10

[0122] A radio signal waveform intelligent sensing system that integrates multi-domain features includes 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 multi-modal features of the signal in different ways; the feature fusion module is used to fuse multi-modal features into more informative fused features through an attention mechanism; the classifier further processes the fused features to obtain the modulation scheme of the final perceived signal.

[0124] The units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0125] According to one aspect of the present invention, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described above.

[0126] In another aspect, embodiments of the present invention also provide a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the methods described in the above embodiments.

Claims

1. A method for intelligent sensing of radio signal waveforms that integrates multi-domain features, characterized in that, include: Feature extraction is performed on the input sequence, and multi-mode features of the signal are extracted in different ways; Then, the multi-modal features are fused into more informative fused features through an attention mechanism; Finally, the fused features are further classified to obtain the modulation scheme of the final perceived signal; The input sequence includes IQ and FFT sequences; The feature extraction of the input sequence, which extracts multi-mode features of the signal in different ways, specifically includes the following sub-steps: performing real convolution (RVC) and complex convolution (CVC) operations on the two sets of sequences respectively to obtain multiple sets of feature streams; The process of fusing multi-modal features into richer fused features through an attention mechanism specifically includes the following sub-steps: The depthwise separable convolutional layer (DSC) is used to split the convolutional layer into two parts, specifically consisting of channel-dimensional convolution and... The convolutional layers transform the dimensions of the output feature map through... This is accomplished through convolution; Furthermore, a channel attention mechanism is adopted to correct the feature strength between channels by modeling the interdependencies between channels. Specifically, this includes: firstly, compressing the feature map and calculating the global average according to the channel dimension; then obtaining the weights corresponding to each channel of the feature map through the activation function; and finally multiplying the feature map with the corresponding weights to obtain the fused features.

2. The intelligent sensing method for radio signal waveforms integrating multi-domain features according to claim 1, characterized in that, The activation function is the Sigmoid activation function.

3. The intelligent sensing method for radio signal waveforms integrating multi-domain features according to claim 1, characterized in that, It also includes a sub-step: a skip connection is provided between the input and output of the attention mechanism to prevent gradient vanishing.

4. The intelligent sensing method for radio signal waveforms integrating multi-domain features according to claim 1, characterized in that, The step of further classifying the fused features to obtain the modulation scheme of the final sensing signal specifically includes the following sub-steps: 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 sensing result.

5. The intelligent sensing method for radio signal waveforms integrating multi-domain features according to claim 4, characterized in that, After receiving the fused features, the residual module and fully connected layer module are used to further process the fused features and map them to the perception result. This process specifically includes the following sub-steps: The residual module consists of two residual units and one pooling layer. Each residual unit consists of two RVC layers and one skip connection; the skip connection is represented as follows: ， in, Indicates the input feature map, This represents the feature map after two RVC operations. This represents the output feature map after the skip connections; during backpropagation, even if the convolution output... For the input feature map The gradient is 0, and the skip connection outputs... Able to guarantee The gradient of the layer is at least 1 to avoid the vanishing gradient phenomenon; while the pooling layer is used to reduce the number of output neurons and prevent overfitting. The fully connected layer module first flattens the feature map into a one-dimensional vector, and then inputs this vector into the fully connected layer, that is: ， in, Indicates the number of fully connected layers. Indicates the first The output of the fully connected layer, and The first Weights and biases in fully connected layers.

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

7. A radio signal waveform intelligent sensing system integrating multi-domain features, characterized in that, It includes a feature extraction module, a feature fusion module, and a classifier; The feature extraction module is used to extract features from the input sequence and extract multi-mode features of the signal in different ways; the feature fusion module is used to fuse multi-mode features into more information-rich fused features through an attention mechanism; the classifier is used to further process the fused features to obtain the modulation mode of the final perceived signal; and is used to execute the intelligent sensing method for radio signal waveforms that fuses multi-domain features as described in claim 1.

Images