Modulation signal recognition method and system for unmanned aerial vehicle communication

Abstract

The present invention relates to the field of automatic modulation classification technology, and specifically discloses a modulation signal recognition and system for unmanned aerial vehicle communication, which is characterized by comprising: step S110, obtaining an original IQ signal; step S120, generating a 4-phase-rotated enhanced signal according to the original IQ signal; step S130, extracting the signal features of the enhanced signal by complex convolution to obtain an input feature map; step S140, using the input feature map as the input of the first layer of lightweight convolution module, and performing lightweight convolution operation and dual-channel attention mechanism operation on the input feature map through cascaded 6-layer lightweight convolution modules to obtain a multi-layer feature map; step S150, performing cross-layer feature fusion operation on the fourth layer feature map, the fifth layer feature map, and the sixth layer feature map in the multi-layer feature map to obtain a modulation signal recognition result. The present invention realizes high recognition accuracy while maintaining the lightweight model, and can be deployed in unmanned aerial vehicle equipment with limited computing resources and storage resources.

Description

Technical Field

[0001] The present invention relates to the field of automatic modulation classification, and in particular to a modulation signal recognition method and system for unmanned aerial vehicle communication. Background Art

[0002] Automatic Modulation Classification (AMC) is a technology that automatically identifies the modulation mode of a signal without prior knowledge or with partial prior knowledge. This process is an intermediate step between signal detection and signal demodulation.

[0003] In the UAV communication system, the application of AMC enables the UAV to adaptively select the most appropriate modulation method according to the current channel conditions and transmission distance, thereby effectively improving the data transmission efficiency. In addition, as a key technology for dynamic spectrum access, AMC plays a significant role in enhancing spectrum efficiency and expanding communication capacity. In a complex electromagnetic environment, AMC technology further demonstrates its potential in monitoring unauthorized devices or signals, especially in the applications of UAV identification, interference identification, and electronic countermeasures. In summary, AMC plays a vital role in optimizing the communication performance of UAVs and ensuring communication security.

[0004] After decades of development, AMC technology has been divided into two categories: traditional recognition technology and deep learning-based recognition technology. Traditional automatic modulation recognition methods can be roughly divided into two categories: methods based on maximum likelihood hypothesis testing and methods based on feature recognition. The method based on maximum likelihood hypothesis testing is rooted in Bayesian theory and aims to achieve the best estimation of the modulation mode by minimizing the probability of misclassification. Although such methods are optimal in theory, their practical application effects are not ideal due to the lack of necessary prior knowledge in non-cooperative communication scenarios. On the other hand, the method based on feature recognition includes two steps: feature extraction and classifier construction. Commonly used features include: constellation features, parameter statistical features, wavelet transform features, etc. These methods usually rely on the experience and knowledge of experts and are limited in the degree of automation. Therefore, with the increasing complexity of communication systems and the need for rapid response in drone systems. Traditional AMC methods can no longer meet current application requirements.

[0005] Deep learning has the ability to automatically extract complex features from training samples and build classifiers, which is very suitable for AMC tasks. However, most of the current AMC methods based on deep learning have many parameters and high computational complexity, and do not fully consider the adaptability of the model to edge devices. In view of this, there is an urgent need for a method that focuses on lightweight and low-complexity models that adapt to edge devices. Summary of the invention

[0006] In view of the above problems, the purpose of the present invention is to provide a modulation signal recognition method and system for UAV communication, which can achieve high-accuracy classification of the modulation signal while keeping the model lightweight.

[0007] The present invention provides a modulation signal recognition method for unmanned aerial vehicle communication, comprising:

[0008] Step S110, obtaining an original IQ signal;

[0009] Step S120, generating a 4-phase rotated enhanced signal according to the original IQ signal;

[0010] Step S130, extracting the signal features of the enhanced signal by complex convolution to obtain an input feature map;

[0011] Step S140, taking the input feature map as the input of the first layer lightweight convolution module, performing lightweight convolution operation and dual-channel attention mechanism operation on the input feature map through the cascaded 6-layer lightweight convolution module to obtain a multi-layer feature map;

[0012] Step S150, performing a cross-layer feature fusion operation on the fourth layer feature map, the fifth layer feature map, and the sixth layer feature map in the multi-layer feature map to obtain a modulation signal recognition result.

[0013] In a possible implementation, step S120 includes:

[0014] Converting the original IQ signal into a discrete time series by sampling;

[0015] The discrete time series is phase rotated around the origin to obtain an enhanced signal.

[0016] In a possible implementation, the expression of the discrete time series r(n) is as follows:

[0017] r(n)=r I (n)+jr Q (n),n=0,1,...,n-1

[0018] The expression of the enhanced signal is as follows:

[0019]

[0020] Among them, r I (n) is the in-phase component, r Q (n) is the orthogonal component, r I '(n) is the real part of the received signal, r Q ′(n) is the imaginary part of the received signal, and θ is the rotation angle.

[0021] In a possible implementation, step S130 includes:

[0022] Extracting features between the in-phase channel and the orthogonal channel components in the enhanced signal by complex convolution;

[0023] Normalizing the features by complex batch normalization to obtain normalized features;

[0024] The normalized features are activated by a complex activation function to obtain the input feature map.

[0025] In a possible implementation, the input feature map M CV The expression is as follows:

[0026] M CV =f CVReLU [f CVBN (w*x)]

[0027] Among them, f CVReLU is the activation function, w is the complex convolution kernel weight, and x is the enhanced signal.

[0028] In a possible implementation, step S140 includes:

[0029] Performing channel-by-channel convolution on each channel of the input feature map;

[0030] Use a (1,1) convolution kernel to perform point-by-point convolution on multiple channels after channel-by-channel convolution, so that the dimension of the combined channels is consistent with the input feature map;

[0031] Rearrange the channels after point-by-point convolution to obtain the intermediate feature map.

[0032] In a possible implementation, in step S140, the first-layer lightweight convolution module performs the following operations: extracting a first aggregated feature in the intermediate feature map through global average pooling and one-dimensional convolution;

[0033] Extracting the first salient feature in the intermediate feature map by global maximum pooling and one-dimensional convolution;

[0034] Vertically splicing the first aggregate feature and the first significant feature along the channel dimension to form a comprehensive channel weight;

[0035] The comprehensive channel weight is multiplied by the intermediate feature map to obtain the first layer feature map.

[0036] In a possible implementation, in step S140, the first-layer lightweight convolution module performs the following operations:

[0037] The output of each layer of lightweight convolution module is used as the input of the next layer of lightweight convolution module, and lightweight convolution operation and dual-channel attention mechanism operation are performed in sequence to obtain the multi-layer feature map.

[0038] In a possible implementation, step S150 includes:

[0039] Performing a global average pooling operation on the fourth layer feature map, the fifth layer feature map, and the sixth layer feature map in parallel, and reducing the dimension by a one-dimensional convolution operation after summing them up, to obtain a second aggregated feature;

[0040] Performing a global maximum pooling operation on the fourth layer feature map, the fifth layer feature map, and the sixth layer feature map in parallel, and reducing the dimension by a one-dimensional convolution operation after summing, to obtain a second significant feature;

[0041] Vertically splicing the second aggregated feature and the second significant feature along the channel dimension to obtain a final feature map;

[0042] The final feature map is connected to the fully connected layer, and the modulation recognition classification result is output after iterative training.

[0043] The present invention also provides a modulation signal recognition system for UAV communication, which is applied to the above-mentioned modulation signal recognition method, comprising:

[0044] An acquisition module, used for acquiring an original IQ signal;

[0045] A phase rotation module, generating a 4-phase rotated enhanced signal according to the original IQ signal;

[0046] A feature extraction module, used to extract the signal features of the enhanced signal through complex convolution to obtain an input feature map;

[0047] A lightweight convolution module, used to take the input feature map as the input of the first layer of lightweight convolution module, and perform lightweight convolution operation and dual-channel attention mechanism operation on the input feature map through cascaded 6 layers of lightweight convolution modules to obtain a multi-layer feature map;

[0048] The cross-layer feature fusion module is used to perform a cross-layer feature fusion operation on the fourth layer feature map, the fifth layer feature map, and the sixth layer feature map in the multi-layer feature map to obtain a modulation signal recognition result.

[0049] The modulation signal recognition method and system for UAV communication provided by the present invention first generates an enhanced signal through the original signal, then designs a lightweight convolution module, a dual-channel attention mechanism module and a cross-layer feature fusion module, and finally trains the network model under the supervision of the loss function to complete the classification of the modulation signal. While maintaining the lightweight of the model, it can also achieve high recognition accuracy and can be deployed on UAV equipment with limited computing resources and storage resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic diagram of a flow chart of a modulation signal identification method provided by an embodiment of the present invention;

[0051] Figure 2 A visualization diagram of phase rotation enhancement of IQ signal sample 4 provided in an embodiment of the present invention;

[0052] Figure 3 A structural diagram of a lightweight convolution module provided in an embodiment of the present invention;

[0053] Figure 4 A structural diagram of a dual-channel attention mechanism module provided by an embodiment of the present invention;

[0054] Figure 5 The overall structure diagram of the lightweight network provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following detailed description of the embodiments of the present invention is further described in detail in conjunction with the accompanying drawings and examples. The detailed description of the following embodiments and the accompanying drawings are used to exemplarily illustrate the principles of the present invention, but cannot be used to limit the scope of the present invention, that is, the present invention is not limited to the preferred embodiments described, and the scope of the present invention is defined by the claims.

[0056] In the description of the present invention, it should be noted that, unless otherwise specified, “plurality” means two or more than two; the terms “first”, “second”, etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance; for ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0057] Figure 1 A flow chart of a modulation signal recognition method provided by an embodiment of the present invention is shown as follows: Figure 1 As shown, the modulation signal recognition method for unmanned aerial vehicle communication provided by the present invention includes:

[0058] Step S110, obtaining an original IQ signal;

[0059] In one possible implementation, the signals received by the drone are usually recorded in IQ mode.

[0060] Step S120, generating a 4-phase-rotated enhanced signal according to the original IQ signal;

[0061] In one possible implementation, the continuous signal is converted into a discrete time series by sampling inside the drone for storage. The discrete time series consists of two data vectors, an in-phase component and an orthogonal component. Specifically, the original IQ signal is converted into a discrete time series by sampling;

[0062] The expression of discrete time series r(n) is as follows:

[0063] r(n)=r I (n)+jr Q (n),n=0,1,...,n-1

[0064] Phase transformation is an effective means of data enhancement in radio signal processing. By rotating the radio signal, data samples with different phase angles can be generated without changing the signal amplitude, thereby increasing the diversity of the data set. Specifically, the discrete time series is phase-rotated around the origin to obtain an enhanced signal.

[0065] The expression of the enhanced signal is as follows:

[0066]

[0067] Among them, r I (n) is the in-phase component, r Q (n) is the orthogonal component, r I '(n) is the real part of the received signal, r Q ′(n) is the imaginary part of the received signal, and θ is the rotation angle.

[0068] Figure 2 The visualization diagrams of the phase rotation enhancement of the IQ signal sample 4 provided in the embodiment of the present invention are schematic diagrams of the received signal being rotated to 0, π / 2, π and 3π / 2 respectively, and the original IQ signal will be amplified four times. Figure 2 The constellation diagram of the CPFSK signal sample is shown in Figure 1. The distribution of sample points obtained after different phase rotations can be intuitively observed.

[0069] Step S130, extracting the signal features of the enhanced signal through complex convolution to obtain an input feature map;

[0070] In a possible implementation, step S130 includes:

[0071] The features between the in-phase channel and the orthogonal channel components in the enhanced signal are extracted through complex convolution;

[0072] The features are normalized by complex batch normalization to obtain normalized features;

[0073] The normalized features are activated by a complex activation function to obtain the input feature map.

[0074] Input feature map M CV The expression is as follows:

[0075] M CV =f CVReLU [f CVBN (w*x)]

[0076] Among them, f CVReLU is the activation function, w is the complex convolution kernel weight, and x is the enhanced signal.

[0077] Step S140, taking the input feature map as the input of the first layer of lightweight convolution module, performing lightweight convolution operation and dual-channel attention mechanism operation on the input feature map through the cascaded 6-layer lightweight convolution module to obtain a multi-layer feature map;

[0078] The input feature map M obtained in step S130 CV Input to the lightweight convolution module LCM for lightweight convolution operation. The LCM module is as follows Figure 3 As shown in . The LCM module mainly consists of separation convolution, channel shuffle and attention mechanism operations. Separation convolution is an efficient convolution operation used to extract feature maps. It mainly includes two parts:

[0079] Depthwise Convolution: Convolution is performed on each channel of the input separately instead of mixing all channels together, which can significantly reduce the amount of calculation and model complexity.

[0080] Pointwise Convolution: After DW, a 1x1 convolution kernel is used to combine channel information in order to enhance the network's expressiveness while keeping the depth of the feature map unchanged.

[0081] After the separation convolution, it is difficult to exchange information between feature maps of different groups, which may lead to reduced network learning efficiency. Therefore, the present invention increases the information flow between output feature maps by adding a channel shuffle operation. Since channel shuffling only achieves information flow by rearranging channels, it avoids the use of additional convolutional or fully connected layers, thereby reducing the number of parameters and computational complexity of the model.

[0082] After shuffling, the output intermediate feature map F is obtained, and the attention mechanism operation is applied to the intermediate feature map F to further realize local cross-channel interaction.

[0083] In one possible implementation, a channel-by-channel convolution is performed on each channel of the input feature map;

[0084] Use a (1,1) convolution kernel to perform point-by-point convolution on multiple channels after channel-by-channel convolution, so that the dimension of the combined channels is consistent with the input feature map;

[0085] Rearrange the channels after point-by-point convolution to obtain the intermediate feature map.

[0086] In one possible implementation, the dual-channel attention mechanism DSECA aims to more effectively highlight useful features and suppress unimportant features, thereby enhancing the model's ability to capture key information. The structure of DSECA is as follows: Figure 4 The specific steps are as follows:

[0087] The first aggregated feature M1 in the intermediate feature map F is extracted through global average pooling GAP and one-dimensional convolution Conv1 D;

[0088] The first significant feature M2 in the intermediate feature map F is extracted through global maximum pooling GMP and one-dimensional convolution Conv1 D;

[0089] The first aggregate feature M1 and the first significant feature M2 are vertically spliced ​​along the channel dimension to form a comprehensive channel weight M;

[0090] Multiply the comprehensive channel weight M by the intermediate feature map F to obtain the first layer feature map F'.

[0091] Step S150, performing a cross-layer feature fusion operation on the fourth layer feature map, the fifth layer feature map, and the sixth layer feature map in the multi-layer feature map to obtain a modulation signal recognition result.

[0092] In one possible implementation, the dual-channel attention mechanism operation DSECA is included in the lightweight convolution module LCM. After the complex convolution, the enhanced signal x is cascaded through 6 layers of LCM modules. The output feature map of the LCM module is defined as F` n , n is the current nth layer LCM module. The cross-layer feature fusion operation acts on the output feature maps F'4, F'5, and F'6 of the last three layers of LCM. Its purpose is to fuse the features of these deep modules. The structure of cross-layer feature fusion is as follows Figure 5 shown.

[0093] Specifically, the fourth-layer feature map, the fifth-layer feature map and the sixth-layer feature map output by the 4th, 5th and 6th-layer lightweight convolution modules are subjected to global average pooling operation G l oba l Average Pool l in parallel, and after summing Add, they are reduced in dimension by one-dimensional convolution operation Conv1 D to obtain the second aggregate feature F Ave ;

[0094] The fourth-layer feature map, the fifth-layer feature map and the sixth-layer feature map output by the 4th, 5th and 6th-layer lightweight convolution modules are subjected to the global maximum pooling operation G l oba l Max Pool l in parallel, and after summing Add, the dimension is reduced by the one-dimensional convolution operation Conv1 D to obtain the second significant feature F Max ;

[0095] The second aggregate feature F Ave and the second significant feature F Max Vertical splicing is performed along the channel dimension to obtain the final feature map F End .

[0096] The final feature map F End Connect the fully connected layer Dense and output the modulation recognition classification results after iterative training.

[0097] The present invention also provides a modulation signal recognition system for UAV communication, which is applied to the above-mentioned modulation signal recognition method, comprising:

[0098] An acquisition module, used for acquiring an original IQ signal;

[0099] A phase rotation module generates a 4-phase rotated enhanced signal according to the original IQ signal;

[0100] A feature extraction module, used for extracting signal features of the enhanced signal through complex convolution to obtain an input feature map;

[0101] A lightweight convolution module is used to take the input feature map as the input of the first layer of lightweight convolution modules, and perform lightweight convolution operations and dual-channel attention mechanism operations on the input feature map through the cascaded 6-layer lightweight convolution modules to obtain a multi-layer feature map;

[0102] The cross-layer feature fusion module is used to perform cross-layer feature fusion operations on the fourth layer feature map, the fifth layer feature map, and the sixth layer feature map in the multi-layer feature map to obtain the modulation signal recognition result.

[0103] The modulation signal recognition method and system for UAV communication provided by the present invention first generates an enhanced signal through the original signal, then designs a lightweight convolution module, a dual-channel attention mechanism module and a cross-layer feature fusion module, and finally trains the network model under the supervision of the loss function to complete the classification of the modulation signal. While maintaining the lightweight of the model, it can also achieve high recognition accuracy and can be deployed on UAV equipment with limited computing resources and storage resources.

[0104] Through experimental verification on the public dataset RML2016.10A, the lightweight network proposed by this invention achieves an average accuracy of 62.83% with only 8815 parameters, and achieves a maximum classification accuracy of 92.11% when the signal-to-noise ratio is 10dB. This shows that the lightweight network proposed by this invention can achieve high recognition accuracy while maintaining the lightweight model, and is very suitable for deployment in drone communication systems.

[0105] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A modulation signal recognition method for unmanned aerial vehicle communication, characterized in that: include: Step S110, obtaining an original IQ signal; Step S120, generating a 4-phase rotated enhanced signal according to the original IQ signal; Step S130, extracting the signal features of the enhanced signal by complex convolution to obtain an input feature map; Step S140, taking the input feature map as the input of the first layer lightweight convolution module, performing lightweight convolution operation and dual-channel attention mechanism operation on the input feature map through the cascaded 6-layer lightweight convolution module to obtain a multi-layer feature map; Step S150, performing a cross-layer feature fusion operation on the fourth layer feature map, the fifth layer feature map, and the sixth layer feature map in the multi-layer feature map to obtain a modulation signal recognition result; In step S140, the first-layer lightweight convolution module performs the following operations: Performing channel-by-channel convolution on each channel of the input feature map; Use a (1,1) convolution kernel to perform point-by-point convolution on multiple channels after channel-by-channel convolution, so that the dimension of the combined channels is consistent with the input feature map; Rearrange the channels after point-by-point convolution to obtain the intermediate feature map; The first-layer lightweight convolution module also performs the following operations: Extracting a first aggregated feature in the intermediate feature map by global average pooling and one-dimensional convolution; Extracting the first salient feature in the intermediate feature map by global maximum pooling and one-dimensional convolution; Vertically splicing the first aggregate feature and the first significant feature along the channel dimension to form a comprehensive channel weight; Multiplying the comprehensive channel weight by the intermediate feature map to obtain a first-layer feature map output by a first-layer lightweight convolution module; The output of each layer of lightweight convolution module is used as the input of the next layer of lightweight convolution module, and lightweight convolution operation and dual-channel attention mechanism operation are performed in sequence to obtain the multi-layer feature map; The step S150 includes: Performing a global average pooling operation on the fourth layer feature map, the fifth layer feature map, and the sixth layer feature map in parallel, and reducing the dimension by a one-dimensional convolution operation after summing them up, to obtain a second aggregated feature; Performing a global maximum pooling operation on the fourth layer feature map, the fifth layer feature map, and the sixth layer feature map in parallel, and reducing the dimension by a one-dimensional convolution operation after summing, to obtain a second significant feature; Vertically splicing the second aggregated feature and the second significant feature along the channel dimension to obtain a final feature map; The final feature map is connected to the fully connected layer, and the modulation recognition classification result is output after iterative training.

2. The modulation signal recognition method according to claim 1, characterized in that: The step S120 includes: Converting the original IQ signal into a discrete time series by sampling; The discrete time series is phase rotated around the origin to obtain an enhanced signal.

3. The modulation signal recognition method according to claim 2, characterized in that: The expression of the discrete time series r(n) is as follows: r(n)=r I (n)+jr Q (n),n=0,1,...,n-1 The expression of the enhanced signal is as follows: Among them, r I (n) is the in-phase component, r Q (n) is the orthogonal component, r′ I (n) is the real part of the received signal, r′ Q (n) is the imaginary part of the received signal, and θ is the rotation angle.

4. The modulation signal recognition method according to claim 1, characterized in that: The step S130 includes: Extracting features between the in-phase channel and the orthogonal channel components in the enhanced signal by complex convolution; Normalizing the features by complex batch normalization to obtain normalized features; The normalized features are activated by a complex activation function to obtain the input feature map.

5. The modulation signal recognition method according to claim 4, characterized in that: The input feature map M CV The expression is as follows: M CV =f CVReLU [f CVBN (w*x)] Among them, f CVReLU is the activation function, w is the complex convolution kernel weight, and x is the enhanced signal.

6. A modulation signal recognition system for drone communication, applied to the modulation signal recognition method according to any one of claims 1 to 5, characterized in that: include: An acquisition module, used for acquiring an original IQ signal; A phase rotation module, generating a 4-phase rotated enhanced signal according to the original IQ signal; A feature extraction module, used to extract the signal features of the enhanced signal through complex convolution to obtain an input feature map; A lightweight convolution module, used to take the input feature map as the input of the first layer of lightweight convolution module, and perform lightweight convolution operation and dual-channel attention mechanism operation on the input feature map through cascaded 6 layers of lightweight convolution modules to obtain a multi-layer feature map; A cross-layer feature fusion module is used to perform a cross-layer feature fusion operation on the fourth layer feature map, the fifth layer feature map, and the sixth layer feature map in the multi-layer feature map to obtain a modulation signal recognition result; The first layer of lightweight convolutional modules performs the following operations: Performing channel-by-channel convolution on each channel of the input feature map; Use a (1,1) convolution kernel to perform point-by-point convolution on multiple channels after channel-by-channel convolution, so that the dimension of the combined channels is consistent with the input feature map; Rearrange the channels after point-by-point convolution to obtain the intermediate feature map; The first-layer lightweight convolution module also performs the following operations: Extracting a first aggregated feature in the intermediate feature map by global average pooling and one-dimensional convolution; Extracting the first salient feature in the intermediate feature map by global maximum pooling and one-dimensional convolution; Vertically splicing the first aggregate feature and the first significant feature along the channel dimension to form a comprehensive channel weight; Multiplying the comprehensive channel weight by the intermediate feature map to obtain a first-layer feature map output by a first-layer lightweight convolution module; The lightweight convolution module is further used to use the output of each layer of lightweight convolution module as the input of the next layer of lightweight convolution module, and perform lightweight convolution operation and dual-channel attention mechanism operation in sequence to obtain the multi-layer feature map; The cross-layer feature fusion module is further used to perform a global average pooling operation on the fourth layer feature map, the fifth layer feature map and the sixth layer feature map in parallel, and reduce the dimension through a one-dimensional convolution operation after summing, so as to obtain a second aggregated feature; Performing a global maximum pooling operation on the fourth layer feature map, the fifth layer feature map, and the sixth layer feature map in parallel, and reducing the dimension by a one-dimensional convolution operation after summing, to obtain a second significant feature; Vertically splicing the second aggregated feature and the second significant feature along the channel dimension to obtain a final feature map; The final feature map is connected to the fully connected layer, and the modulation recognition classification result is output after iterative training.

Images