Unmanned aerial vehicle signal modulation identification method and system based on multistage attention mechanism

Abstract

The invention relates to an unmanned aerial vehicle signal modulation identification method and system based on a multistage attention mechanism, and the method comprises the steps: S1, obtaining the image transmission signal data of an unmanned aerial vehicle, and carrying out the preprocessing of the image transmission signal data of the unmanned aerial vehicle, and obtaining an input feature map; s2, constructing an unmanned aerial vehicle signal modulation recognition model, and inputting the input feature map into the unmanned aerial vehicle signal modulation recognition model for training; and S3, performing unmanned aerial vehicle signal identification by using the trained unmanned aerial vehicle signal modulation identification model. According to the method, a channel attention mechanism is adopted, different channel weights are output through extrusion excitation operation, so that the model pays attention to important features and ignores secondary features, and dynamic adjustment of each channel weight is achieved; a self-attention mechanism is adopted to pay attention to global information of the input feature map. The method still has high identification accuracy in a low signal-to-noise ratio environment, and compared with ResNet50, SeNet, CNN, DNN and other networks, the method has the advantages of high identification accuracy and strong robustness.

Description

Technical Field

[0001] The present invention relates to a method and system for identifying modulation of unmanned aerial vehicle (UAV) signals based on a multi-level attention mechanism, and belongs to the field of UAV communication and signal processing. Background Art

[0002] In recent years, the low-altitude intelligent internet has rapidly developed. The low-altitude communication network, low-altitude perception network, and other digital networks within this low-altitude intelligent internet, as well as services such as express logistics, inspections, and emergency rescue, all require drones as carriers. However, with the substantial increase in the number of commercial and civilian drones, the social security issues brought about by the phenomenon of "illegal flying" (such as disrupting flights, leaking privacy, and threatening personal safety) cannot be ignored. Therefore, how to regulate and identify drones is a key issue that urgently needs to be addressed.

[0003] Drone signals primarily consist of remote control signals and image transmission signals. Mainstream drone remote control signals typically operate in the 2.4 GHz frequency band and utilize frequency hopping communication, while drone image transmission signals primarily operate in the 5.8 GHz frequency band and utilize non-frequency hopping communication. Identifying the modulation scheme of drone image transmission signals can be used to screen drones. Traditional modulation scheme identification methods rely on statistical signal characteristics, such as analyzing the signal envelope and instantaneous phase difference, but are difficult to effectively identify in low signal-to-noise ratio (SNR) and complex environments. In recent years, machine learning has made progress in multiple fields due to its unique advantages. In modulation scheme identification, some researchers have used convolutional neural networks to extract signal features and classify modulated signals. However, these methods also suffer from poor algorithm robustness and high computational complexity.

[0004] The attention mechanism mimics the different levels of attention humans pay attention to when dealing with problems, filtering out important information from a large amount of information while ignoring secondary information. In the problem of identifying the modulation of drone image transmission signals, the attention mechanism can accurately capture the dependencies between signals, extract the main features of the signals, and complete the efficient identification of the modulation mode of drone image transmission signals in low signal-to-noise ratio and dynamic environments. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the present invention provides a method and system for UAV signal modulation recognition based on a multi-level attention mechanism, including UAV transmission signal, transmission channel, receiving processing, and signal recognition; the UAV transmission signal is obtained by matlab simulation, the transmission channel is an AWGN channel, the receiving processing is the receiving signal, and the signal recognition process includes data preprocessing, feature extraction, and signal classification; the UAV transmission signal is recorded as , the signal received by the receiving end is recorded as , , represents the wireless channel impulse response, represents additive white Gaussian noise; the received signal is processed by data preprocessing to suppress unimportant features and highlight important features, and then becomes a modulation mode that can be recognized by the neural network. The calculation process is expressed as , represents the extracted features, Represents the signal preprocessing function, and finally the extracted features are sent to the neural network to identify the signal modulation mode. , represents the neural network classification function, Indicates the debugging mode identification result.

[0006] Explanation of terms: 1. SE: It is the abbreviation of Squeeze Excitation. Squeeze excitation adaptively adjusts the feature weights of each channel, suppresses unimportant channel information, and enhances the expression ability of key channel information.

[0007] 2. SA: It is the abbreviation of Self Attention, which is the self-attention mechanism; it captures the dependencies between different elements in the sequence regardless of distance and outputs an adaptive weight matrix.

[0008] 3. Activation function: Common activation functions include Softmax function, Sigmoid function, ReLu function, Tabh function, etc., which are important methods for introducing nonlinearity into neural networks. Softmax function is often used in the output layer of multi-classification problems.

[0009] The technical solutions of the present invention are as follows: A first aspect of the present invention provides a method for identifying modulation of drone signals based on a multi-level attention mechanism, comprising: S1. Obtain the drone image transmission signal data, pre-process the drone image transmission signal data, and obtain the input feature map; S2. Build a UAV signal modulation recognition model and input the input feature map into the UAV signal modulation recognition model for training; S3. Use the trained drone signal modulation recognition model to perform drone signal recognition.

[0010] According to a preferred embodiment of the present invention, in step S1, the UAV image transmission signal data is obtained, and the UAV image transmission signal data is preprocessed to obtain an input feature map; including: Relying on Matlab R2020a as the data generation platform, the drone image transmission signals of eight digital modulation modes, including BPSK, QPSK, 8PSK, BFSK, QFSK, 4QAM, 64QAM, and QASK, were simulated and generated to obtain the drone image transmission signals. Additive white Gaussian noise was added to the drone image transmission signals, and signal samples were randomly generated within the signal-to-noise ratio range. The signal samples were then divided into training, validation, and test sets. To improve the recognition accuracy of the model, we first need to preprocess the drone image transmission signal and decompose the received signal, i.e., the drone image transmission signal, to obtain the input feature map. This includes: The received signal is a complex signal, and the neural network can only process real data, so the complex data is converted to the real domain, and the received signal is , decompose the complex signal as shown below: in, represents the matrix composed of the real parts of the signal elements, represents the matrix composed of the imaginary parts of the signal elements, Indicates the signal length.

[0011] Preferably, according to the present invention, a UAV signal modulation recognition model is constructed, and the input feature map is input into the UAV signal modulation recognition model for training; comprising: The UAV signal modulation recognition model includes: a channel attention mechanism model based on the squeeze excitation module and a self-attention mechanism; The channel attention mechanism model based on the squeeze-excitation module includes: convolution module and squeeze-excitation module; The convolution module includes: the first convolution layer, the first batch of normalization layers, the second convolution layer, and the second batch of normalization layers; The squeeze incentive module includes: squeeze operation and incentive operation; Preferably, according to the present invention, the trained UAV signal modulation recognition model is used to perform UAV signal recognition, including: The input feature map is input into the channel attention mechanism model based on the squeeze excitation module for feature extraction. The channel attention mechanism outputs different channel weights so that the model focuses on important features and ignores minor features. The input feature map is input into the convolution module, and the number of output channels is 16 and the convolution kernel size is , the step size is 1, the padding method is the same as the first convolution layer for convolution operation, and then the output of the first convolution layer is processed by the first batch of normalization layers to accelerate the training process of the network, standardize each channel of the feature map, and then the number of output channels is 32 and the convolution kernel size is , the step size is 1, the filling method is the same as the second convolution layer, the convolution operation and the second batch of normalization layer processing are performed to obtain the output feature map ; Small convolution kernel helps to speed up network training and reduce the number of parameters. High number of output channels maps data from low-dimensional to high-dimensional space, and the input feature map is recorded as , the output feature map is recorded as ,in, Indicates the number of channels, represents the input feature map height, represents the width of the input feature map, Indicates the number of input feature map channels, represents the output feature map height, represents the output feature map width, Indicates the number of output feature map channels, represents the field of real numbers; Output feature map Perform a squeeze operation, including outputting feature maps After the global average pooling layer, the information of each channel is compressed into a scalar to obtain the global features of each channel , as shown below: in, , Output feature map Middle The weight of each channel, For extrusion operation, For channel In position The eigenvalue at , this operation can obtain the global features of each channel ; Global features for each channel Perform excitation operations, including the global features of each channel The data is fed into two Dense fully connected layers, namely the first fully connected layer and the second fully connected layer. The number of output nodes of the first fully connected layer and the second fully connected layer are 64 and 32 respectively. The weight coefficient of each channel is extracted. , as shown below: in, , is the weight of the c-th channel, For the fully connected layer operation, represents the first fully connected layer, Represents the second fully connected layer. Both the first and second fully connected layers use the ReLu activation function to accelerate network training and reduce the gradient vanishing problem. The weight coefficient of each channel Output feature map before squeeze excitation operation Multiply them so that each channel is assigned a different weight, as shown in the following formula: in, Indicates the The output feature map of the channels, Represents the feature map after redistribution of weights. The larger the value, the more important the channel. Feature maps that will be re-weighted Output feature map before squeeze excitation operation Adding preserves the information of the original feature map and helps alleviate the gradient disappearance problem and improve the network training process, as shown in the following formula: in, Indicates the The output feature map after the channels are re-assigned weights is Perform the above formula (III) (IV) (V) operation again, further assign different weights to each channel, and obtain the output feature map ;Finally, the self-attention mechanism is used to output the feature map The features are further extracted, and after multiple rounds of iterative training, a trained drone signal modulation recognition model is obtained.

[0012] According to the present invention, the self-attention mechanism is preferably used to output the feature map Further feature extraction; including: The self-attention mechanism can focus on the global information of the input feature map and is independent of the order of the feature map, so that the relationship between elements in different positions can be represented. Using multiple different attention heads can focus on different relationships in the input feature map. The multi-head attention mechanism is the concatenation of the output results of multiple different attention heads. The output feature map of the squeeze excitation module , respectively, with the output feature map The weight matrix is multiplied to generate the query Q (Query), key K (Key) and value V (Value), as shown in the following formula: in, , , They are query Q , key K Sum V The weight matrix, represents the field of real numbers, Indicates the number of channels of the output feature map D, represents the height of the output feature map D, Represents the width of the output feature map D; Reshape the query Q, key K and value V to get , calculation query With all keys The dot product of is used to obtain the relevance of each key to the query, and then these relevances are converted into weights through the softmax operation, as shown in the following formula: in, represents the first attention head, query Indicates the degree of attention of the current element to other elements. Indicates the response of all elements to the query, the value Represents the actual information of all elements, the dimension of the key vector , used to scale the dot product to avoid gradient vanishing due to excessive values; The second attention head The calculation is the same as the formula , concatenating the outputs of the two attention heads according to the last dimension yields the result of the multi-head attention or self-attention mechanism, as shown in the following formula: Among them, the spliced attention head It is a high-dimensional tensor; after the flattening operation, the high-dimensional tensor is flattened into a one-dimensional vector, and finally the final result of the modulation method is obtained through the fully connected layer, as shown in the following formula: in, Indicates that The flattened one-dimensional vector, Represents a flattening operation, is the number of modulation types, is the prediction result output by the fully connected layer, Indicates the calculation method of the activation function, The calculation is as follows: in, For the function The output of a neuron, For input data, is the i-th element in the input vector, representing the raw scores (logits) of the i-th category, which are the model’s predicted output for the input data; is the jth element in the input vector, representing the raw score of the jth category. In the denominator of the Softmax function, the raw scores of all n categories are summed, where j is an index from 1 to n. is the number of categories; go through After the activation function is processed, the output result is the probability distribution of eight modulation mode categories. The probability value of each category is between 0 and 1, and the sum of the probabilities of all categories is 1. The maximum probability is selected as the prediction result of the category.

[0013] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of a method for identifying modulation of drone signals based on a multi-level attention mechanism are implemented.

[0014] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for identifying modulation of drone signals based on a multi-level attention mechanism.

[0015] A second aspect of the present invention provides a drone signal modulation recognition system based on a multi-level attention mechanism, comprising: The preprocessing module is configured to: obtain the UAV image transmission signal data, preprocess the UAV image transmission signal data, and obtain an input feature map; The model training module is configured to: construct a UAV signal modulation recognition model and input the input feature map into the UAV signal modulation recognition model for training; The prediction module is configured to use the trained drone signal modulation recognition model to perform drone signal recognition.

[0016] The beneficial effects of the present invention are: 1. The present invention decomposes the input complex data into two channels of real data, and uses the noisy data as network input after passing through the additive white Gaussian noise channel, setting different signal-to-noise ratios to simulate various channel environments.

[0017] 2. This invention uses a channel attention mechanism to output different channel weights through a squeeze-excitation operation, so that the model focuses on important features and ignores minor features, achieving dynamic adjustment of each channel weight. This invention uses a channel attention mechanism to output different channel weights through a squeeze-excitation operation, so that the model focuses on important features and ignores minor features, achieving dynamic adjustment of each channel weight.

[0018] 3. The present invention adopts a self-attention mechanism to focus on the global information of the input feature map, and is independent of the order of the feature map, so that the relationship between elements in different positions can be represented. By using multiple different attention heads, different relationships of the input feature map can be focused on. Splicing the output results of multiple different attention heads is a multi-head attention mechanism.

[0019] 4. The softmax activation function of the present invention solves the multi-classification problem of the output layer. The output result is the probability distribution of the category. The probability value of each category is between 0 and 1, and the sum of the probabilities of all categories is 1. The maximum probability is selected as the prediction result of the category, which can achieve a high-accuracy classification result.

[0020] 5. The present invention still has a high recognition accuracy in a low signal-to-noise ratio environment. Compared with ResNet50, SeNet, CNN, DNN and other networks, it has high recognition accuracy and strong robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a schematic diagram of a UAV image transmission signal transmission and reception model of the UAV signal modulation recognition method based on a multi-level attention mechanism of the present invention; Figure 2 This is a schematic diagram of the modulation recognition accuracy of the model proposed in the present invention when using different numbers of signal samples under different signal-to-noise ratios; Figure 3 This is a schematic diagram of the specific steps of signal modulation mode identification of the model proposed in the present invention; Figure 4 This is the architecture diagram of the channel attention mechanism model based on the squeeze incentive module of the model proposed in this invention; Figure 5 This is a comparison chart of the ablation experiment of the model proposed in the present invention; Figure 6 This is a schematic diagram of the overall recognition accuracy of the model proposed in this invention under different signal-to-noise ratios; Figure 7 This is a schematic diagram of the recognition accuracy of each modulation mode under different signal-to-noise ratios of the model proposed in the present invention; Figure 8 This is a comparison chart of the recognition accuracy of the model proposed in this invention and other classic network modulation methods. DETAILED DESCRIPTION

[0022] The present invention will be further described below with reference to embodiments and accompanying drawings, but is not limited thereto.

[0023] Example 1 A modulation recognition method for drone signals based on a multi-level attention mechanism, such as Figure 1 Shown, including: S1. Obtain the drone image transmission signal data, pre-process the drone image transmission signal data, and obtain the input feature map; S2. Build a UAV signal modulation recognition model and input the input feature map into the UAV signal modulation recognition model for training; S3. Use the trained drone signal modulation recognition model to perform drone signal recognition.

[0024] Example 2 The difference between the method for identifying modulation of drone signals based on a multi-level attention mechanism described in Example 1 is that: In step S1, the UAV image transmission signal data is obtained and preprocessed to obtain an input feature map; including: Relying on Matlab R2020a as the data generation platform, the drone image transmission signals of eight digital modulation modes including BPSK, QPSK, 8PSK, BFSK, QFSK, 4QAM, 64QAM and QASK were simulated and generated to obtain the drone image transmission signals. Among them, BPSK (Binary Phase Shift Keying) represents binary data (0 and 1) through two phases (0° and 180°), transmitting 1 bit per symbol. It has strong noise resistance and simple implementation, but the spectrum efficiency is low (1 bit / Hz). Quadrature PSK (QPSK) uses four phases (e.g., 0°, 90°, 180°, and 270°) to represent two bits of data (00, 01, 10, and 11). Each symbol transmits two bits, with a spectral efficiency twice that of BPSK (2 bit / Hz), but requires a higher signal-to-noise ratio (SNR) and has weaker immunity to phase noise. 8PSK (8-ary Phase Shift Keying) uses eight phases (spaced 45° apart) to represent three bits of data. Each symbol transmits three bits, with higher spectral efficiency (3 bit / Hz), but is sensitive to the SNR and phase noise. BFSK (Binary Frequency Shift Keying) uses two carrier frequencies of different frequencies to represent 0 and 1, respectively. Each symbol transmits one bit, with strong immunity to amplitude interference, but occupies a larger bandwidth and has lower spectral efficiency (1 bit / Hz). Quaternary frequency shift keying (QFSK) uses four carrier waves of different frequencies to represent 2-bit data (e.g., 00, 01, 10, 11). With 2 bits per symbol, the spectrum efficiency is higher than that of BFSK (2 bits / Hz), but the frequency spacing must be properly designed to avoid interference. 4QAM (4-ary Quadrature Amplitude Modulation) simultaneously modulates the amplitude and phase of quadrature carrier waves (both I and Q) to form four constellation points. 64QAM (64-ary Quadrature Amplitude Modulation) combines 16 amplitude and phase points to form 64 constellation points, transmitting 6 bits of data per symbol. It has extremely high spectrum efficiency (6 bits / Hz), but requires an extremely high signal-to-noise ratio and is sensitive to channel distortion and noise. 4ASK (4-ary Amplitude Shift Keying) uses four different amplitude levels to represent 2-bit data (e.g., 00, 01, 10, 11); 2 bits per symbol, simple to implement, but the amplitude is easily affected by channel attenuation and interference; Add additive Gaussian white noise to the drone's image transmission signal, randomly generate signal samples within the signal-to-noise ratio range, and divide the signal samples into training set, validation set, and test set; the signal-to-noise ratio range is -20dB to 10dB, with an interval of 2dB, such as Figure 2 As shown in the figure, the effect of different numbers of signal samples on the network recognition performance was tested. When the number of signal samples was 800, 1000, and 2000, the modulation recognition accuracy was low. When the number of signal samples was 5000, 6000, and 7000, the modulation recognition accuracy was high and the recognition performance was close. Therefore, 5000 signal samples were randomly generated within the signal-to-noise ratio range for each modulation mode. The total number of samples was 5000×8=40000, 90% of which were used as training sets and 10% as validation sets. The signal-to-noise ratio (SNR) of the test set ranges from -20dB to 10dB, with a 2dB interval. 1000 signal samples are generated for each modulation mode signal at each SNR, for a total of 1000 × 8 × 16 = 128,000 samples. The test set is fed into the trained network for recognition, and the network's recognition accuracy under different SNR conditions can be determined. The specific steps of signal recognition are as follows: Figure 3 As shown in the figure, in order to improve the recognition accuracy of the model, it is necessary to first preprocess the drone image transmission signal and decompose the received signal, i.e., the drone image transmission signal, to obtain the input feature map; this includes: The received signal is a complex signal, and the neural network can only process real data, so the complex data is converted to the real domain, and the received signal is , decompose the complex signal as shown below: in, represents the matrix composed of the real parts of the signal elements, represents the matrix composed of the imaginary parts of the signal elements, Indicates the signal length.

[0025] Construct a UAV signal modulation recognition model and input the input feature map into the UAV signal modulation recognition model for training; including: The UAV signal modulation recognition model includes: a channel attention mechanism model based on the squeeze excitation module and a self-attention mechanism; The channel attention mechanism model based on the squeeze-excitation module includes: convolution module and squeeze-excitation module; The convolution module includes: the first convolution layer, the first batch of normalization layers, the second convolution layer, and the second batch of normalization layers; The squeeze incentive module includes: squeeze operation and incentive operation; Use the trained drone signal modulation recognition model to identify drone signals; including: The input feature map is input into the channel attention mechanism model based on the squeeze-excitation module for feature extraction; Figure 4 As shown, the channel attention mechanism outputs different channel weights so that the model focuses on important features and ignores minor features; the input feature map is input into the convolution module, the number of output channels is 16, and the convolution kernel size is , the step size is 1, the padding method is the same as the first convolution layer for convolution operation, and then the output of the first convolution layer is processed by the first batch of normalization layers to accelerate the training process of the network, standardize each channel of the feature map, and then the number of output channels is 32 and the convolution kernel size is , the step size is 1, the filling method is the same as the second convolution layer, the convolution operation and the second batch of normalization layer processing are performed to obtain the output feature map ; Small convolution kernel helps to speed up network training and reduce the number of parameters. High number of output channels maps data from low-dimensional to high-dimensional space, and the input feature map is recorded as , the output feature map is recorded as ,in, Indicates the number of channels, represents the input feature map height, represents the width of the input feature map, Indicates the number of input feature map channels, represents the output feature map height, represents the output feature map width, Indicates the number of output feature map channels, represents the field of real numbers; Output feature map Perform a squeeze operation, including outputting feature maps After the global average pooling layer, the information of each channel is compressed into a scalar to obtain the global features of each channel , as shown below: in, , Output feature map Middle The weight of each channel, For extrusion operation, For channel In position The eigenvalue at this location can be used to obtain the global features of each channel. ; Global features for each channel Perform excitation operations, including the global features of each channel The data is fed into two Dense fully connected layers, namely the first fully connected layer and the second fully connected layer. The number of output nodes of the first fully connected layer and the second fully connected layer are 64 and 32 respectively. The weight coefficient of each channel is extracted. , as shown below: in, , is the weight of the c-th channel, For the fully connected layer operation, represents the first fully connected layer, Represents the second fully connected layer. Both the first and second fully connected layers use the ReLu activation function to accelerate network training and reduce the gradient vanishing problem. The weight coefficient of each channel Output feature map before squeeze excitation operation Multiply them so that each channel is assigned a different weight, as shown in the following formula: in, Indicates the The output feature map of the channels, Represents the feature map after redistribution of weights. The larger the value, the more important the channel. Feature maps to which weights will be redistributed Output feature map before squeeze excitation operation Adding preserves the information of the original feature map and helps alleviate the gradient disappearance problem and improve the network training process, as shown in the following formula: in, Indicates the The output feature map after the channels are re-assigned weights is Perform the above formula (III) (IV) (V) operation again, further assign different weights to each channel, and obtain the output feature map ;Finally, the self-attention mechanism is used to output the feature map The features are further extracted, and after multiple rounds of iterative training, a trained drone signal modulation recognition model is obtained.

[0026] Use the self-attention mechanism to output feature maps Further feature extraction; including: The self-attention mechanism can focus on the global information of the input feature map and is independent of the order of the feature map, so that the relationship between elements in different positions can be represented. Using multiple different attention heads can focus on different relationships in the input feature map. The multi-head attention mechanism is the concatenation of the output results of multiple different attention heads. The output feature map of the squeeze excitation module , respectively, with the output feature map The weight matrix is multiplied to generate the query Q (Query), key K (Key) and value V (Value), as shown in the following formula: in, , , They are query Q , key K Sum V The weight matrix, represents the field of real numbers, Indicates the number of channels of the output feature map D, represents the height of the output feature map D, Represents the width of the output feature map D; Reshape the query Q, key K and value V to get , calculation query With all keys The dot product of is used to obtain the relevance of each key to the query, and then these relevances are converted into weights through the softmax operation, as shown in the following formula: in, represents the first attention head, query Indicates the degree of attention of the current element to other elements. Indicates the response of all elements to the query, value Represents the actual information of all elements, the dimension of the key vector , used to scale the dot product to avoid gradient vanishing due to excessive values; The second attention head The calculation is the same as the formula , concatenating the outputs of the two attention heads according to the last dimension yields the result of the multi-head attention or self-attention mechanism, as shown in the following formula: Among them, the spliced attention head It is a high-dimensional tensor. After the flattening operation, the high-dimensional tensor is flattened into a one-dimensional vector. Finally, the full connection layer is passed to obtain the final result of the modulation method, as shown in the following formula: in, Indicates that The flattened one-dimensional vector, Represents a flattening operation, is the number of modulation types, is the prediction result output by the fully connected layer, Indicates the calculation method of the activation function, The calculation is as follows: in, For the function The output of a neuron, For input data, is the i-th element in the input vector, representing the raw scores (logits) of the i-th category, which are the model’s predicted output for the input data; is the jth element in the input vector, representing the raw score of the jth category. In the denominator of the Softmax function, the raw scores of all n categories are summed, where j is an index from 1 to n. is the number of categories; go through After the activation function is processed, the output result is the probability distribution of eight modulation modes. The probability value of each category is between 0 and 1, and the sum of the probabilities of all categories is 1. The maximum probability is selected as the predicted result of the category (the predicted result is one of the eight modulation modes, and the one with the maximum probability is selected), which can achieve a high-accuracy classification result.

[0027] In order to illustrate the impact of each module on network performance, we removed the channel attention module, the self-attention mechanism, and removed both modules at the same time to conduct ablation experiments and compare them, which are respectively denoted as SA (without SE), SE (without SA), and CNN (without SA and SE). Figure 5 As shown, the channel attention module and self-attention mechanism both have an impact on network performance improvement. When the signal-to-noise ratio is 0dB or below, the network recognition accuracy is low. As the signal-to-noise ratio increases, the recognition accuracy of each network improves. CNN improves the slowest and eventually stabilizes at 92%, indicating that the single CNN network has poor feature extraction capabilities. The SE and SA modules improve performance and can both achieve recognition accuracy of over 98% at 18dB. The SESA model proposed in this invention, i.e., the drone signal modulation recognition model, has the fastest recognition accuracy improvement when the signal-to-noise ratio is 0dB or above, reaching 100% recognition accuracy at 10dB. This shows that the combination of the two modules can greatly improve network recognition performance.

[0028] Figure 6 and Figure 7These are the overall accuracy of the model's signal modulation recognition when the signal-to-noise ratio (SNR) ranges from -20dB to 10dB, and the recognition accuracy of each modulation mode. QASK has a distinct recognition feature and can be accurately recognized at -15dB, while the recognition accuracy of other modulation modes is less than 30% at -15dB. As the SNR increases, the recognition accuracy of each modulation mode increases and reaches over 95% at 10dB. The model proposed in this invention is highly robust and can maintain a high recognition accuracy even in low SNR environments, making it applicable to complex channel environments.

[0029] The SESA model proposed in this invention is compared with the classic network model in terms of modulation recognition accuracy. Figure 8 As shown. It can be seen from the figure that with the increase of the signal-to-noise ratio, the recognition accuracy of all network models is improving. The full connectivity of the DNN network and the lightweight nature of MobileNet lead to lower recognition accuracy. SENet is close to the performance of the proposed model at 5dB and below. As the signal-to-noise ratio increases, the recognition accuracy decreases. ResNet50 has a higher recognition accuracy as the signal-to-noise ratio increases, but it has the disadvantages of a complex model and a large amount of calculation. The SESA model proposed in the present invention shows the best network performance in all signal-to-noise ratio ranges. The model makes full use of the dynamic adjustment of channel weights of the channel attention mechanism and the global information dependency of the self-attention mechanism, and can extract data features to the maximum extent to achieve high recognition accuracy.

[0030] Example 3 A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of the drone signal modulation recognition method based on the multi-level attention mechanism described in Example 1 or 2.

[0031] Example 4 A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the drone signal modulation recognition method based on the multi-level attention mechanism described in Example 1 or 2 are implemented.

[0032] Example 5 A drone signal modulation recognition system based on a multi-level attention mechanism, including: The preprocessing module is configured to: obtain the UAV image transmission signal data, preprocess the UAV image transmission signal data, and obtain an input feature map; The model training module is configured to: construct a UAV signal modulation recognition model and input the input feature map into the UAV signal modulation recognition model for training; The prediction module is configured to use the trained drone signal modulation recognition model to perform drone signal recognition.

Claims

1. A method for UAV signal modulation recognition based on a multi-level attention mechanism, characterized in that: include: S1. Obtain the drone image transmission signal data, pre-process the drone image transmission signal data, and obtain the input feature map; S2. Build a UAV signal modulation recognition model and input the input feature map into the UAV signal modulation recognition model for training; S3. Use the trained drone signal modulation recognition model to perform drone signal recognition.

2. The method for drone signal modulation recognition based on a multi-level attention mechanism according to claim 1 is characterized in that: In step S1, the UAV image transmission signal data is obtained and preprocessed to obtain an input feature map; including: Relying on Matlab R2020a as the data generation platform, the drone image transmission signals of 8 digital modulation modes including BPSK, QPSK, 8PSK, BFSK, QFSK, 4QAM, 64QAM and QASK are simulated and generated to obtain the drone image transmission signals; Add additive Gaussian white noise to the drone's image transmission signal, randomly generate signal samples within the signal-to-noise ratio range, and divide the signal samples into training set, validation set, and test set; Decompose the received signal, i.e., the drone image transmission signal, to obtain the input feature map; including: The received signal is a complex signal. Convert the complex signal data to the real number domain. The received signal is , decompose the complex signal as shown below: in, represents the matrix composed of the real parts of the signal elements, represents the matrix composed of the imaginary parts of the signal elements, Indicates the signal length.

3. The method for drone signal modulation recognition based on a multi-level attention mechanism according to claim 1 is characterized in that: Construct a UAV signal modulation recognition model and input the input feature map into the UAV signal modulation recognition model for training; including: The UAV signal modulation recognition model includes: a channel attention mechanism model based on the squeeze excitation module and a self-attention mechanism; The channel attention mechanism model based on the squeeze-excitation module includes: convolution module and squeeze-excitation module; The convolution module includes: the first convolution layer, the first batch of normalization layers, the second convolution layer, and the second batch of normalization layers; The squeeze incentive module includes: squeeze operation and incentive operation.

4. The method for drone signal modulation recognition based on a multi-level attention mechanism according to claim 3 is characterized in that: Use the trained drone signal modulation recognition model to identify drone signals; including: The input feature map is input into the channel attention mechanism model based on the squeeze excitation module for feature extraction, including inputting the input feature map into the convolution module, performing convolution operation through the first convolution layer, then performing the first batch of normalization layer processing on the output of the first convolution layer, and then performing convolution operation and the second batch of normalization layer processing on the output of the first convolution layer to obtain the output feature map. ; The input feature map is recorded as , the output feature map is recorded as ,in, Indicates the number of channels, represents the input feature map height, represents the width of the input feature map, Indicates the number of input feature map channels, represents the output feature map height, represents the output feature map width, Indicates the number of output feature map channels, represents the field of real numbers; Output feature map Perform a squeeze operation, including outputting feature maps After the global average pooling layer, the global features of each channel are obtained , as shown below: in, , Output feature map Middle The weight of each channel, For extrusion operation, For channel In position The eigenvalue at ; Global features for each channel Perform excitation operations, including the global features of each channel The weight coefficients of each channel are extracted by sequentially sending them to two Dense fully connected layers, namely the first fully connected layer and the second fully connected layer. , as shown below: in, , is the weight of the c-th channel, For the fully connected layer operation, represents the first fully connected layer, Represents the second fully connected layer. Both the first and second fully connected layers use the ReLu activation function. The weight coefficient of each channel Output feature map before squeeze excitation operation Multiply them so that each channel is assigned a different weight, as shown in the following formula: in, Indicates the The output feature map of the channels, Represents the feature map after redistribution of weights; Feature maps that will be re-weighted Output feature map before squeeze excitation operation Add them together, as shown below: in, Indicates the The output feature map after the channels are re-assigned weights is Perform the above formula (III) (IV) (V) operation again to obtain the output feature map ;Finally, the self-attention mechanism is used to output the feature map The features are further extracted, and after multiple rounds of iterative training, a trained drone signal modulation recognition model is obtained.

5. The method for drone signal modulation recognition based on a multi-level attention mechanism according to claim 4 is characterized in that: Use the self-attention mechanism to output feature maps Further feature extraction; including: The output feature map of the squeeze excitation module , respectively, with the output feature map The weight matrix is multiplied to generate the query Q (Query), key K (Key) and value V (Value), as shown in the following formula: in, , , They are query Q , key K Sum V The weight matrix, represents the field of real numbers, Indicates the number of channels of the output feature map D, represents the height of the output feature map D, Represents the width of the output feature map D; Reshape the query Q, key K and value V to get , calculation query With all keys The dot product of is used to obtain the relevance of each key to the query, and then these relevances are converted into weights through the softmax operation, as shown in the following formula: in, represents the first attention head, query Indicates the degree of attention of the current element to other elements. Indicates the response of all elements to the query, the value Represents the actual information of all elements, the dimension of the key vector , used to scale the dot product; The second attention head The calculation is the same as the formula , concatenating the outputs of the two attention heads according to the last dimension yields the result of the multi-head attention or self-attention mechanism, as shown in the following formula: Among them, the spliced attention head It is a high-dimensional tensor; after the flattening operation, the high-dimensional tensor is flattened into a one-dimensional vector, and finally the final result of the modulation method is obtained through the fully connected layer, as shown in the following formula: in, Indicates that The flattened one-dimensional vector, Represents a flattening operation, is the number of modulation types, is the prediction result output by the fully connected layer, Indicates the calculation method of the activation function; go through After the activation function is processed, the output result is the probability distribution of eight modulation mode categories. The probability value of each category is between 0 and 1, and the sum of the probabilities of all categories is 1. The maximum probability is selected as the prediction result of the category.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of the drone signal modulation recognition method based on a multi-level attention mechanism described in any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the drone signal modulation recognition method based on a multi-level attention mechanism described in any one of claims 1 to 5 are implemented.

8. A UAV signal modulation recognition system based on a multi-level attention mechanism, characterized by: include: The preprocessing module is configured to: obtain the UAV image transmission signal data, preprocess the UAV image transmission signal data, and obtain an input feature map; The model training module is configured to: construct a UAV signal modulation recognition model and input the input feature map into the UAV signal modulation recognition model for training; The prediction module is configured to use the trained drone signal modulation recognition model to perform drone signal recognition.

Images