An efficient method and device for identifying radar signal modulation modes

Abstract

The present invention discloses an efficient method and device for identifying radar signal modulation modes. Specifically, a spatio-temporal feature sharing reconstruction network STARNet is proposed to achieve excellent recognition accuracy in a lightweight manner. A single autoencoder structure is used to simultaneously extract the low-dimensional spatial and temporal features of radio signals. In addition, a hybrid attention module (HA-Ghost) is designed to automatically extract discriminative radar signal spatial information according to the signal reconstruction performance, thereby reducing the number of model parameters and improving the performance of automatic recognition. A large number of experiments on benchmark datasets show that the proposed STARNet achieves an average modulation classification accuracy of 63.60%, which is better than the previous state-of-the-art models. Moreover, although more types of features are extracted, STARNet has only 14,860 parameters, which is much less than the parameters of existing autoencoder-based methods.

Description

Technical Field

[0001] The present invention belongs to the fields of pattern recognition and information technology and can be used for radar signal recognition. More specifically, the present invention relates to a method and device for efficiently recognizing radar signal modulation modes. Background Art

[0002] Automatic modulation classification of radar signals is the process of identifying the modulation mode of received radar signals without knowing the radar communication system, and it is an important part of non-cooperative communication in the field of wireless communication. With the remarkable development of wireless communication technology, the types of modulation modes have become more diverse, and the number of wireless devices has increased rapidly, resulting in an increasingly complex communication environment. Identifying various radar signal modulation modes with high precision in a low signal-to-noise ratio environment while maintaining computational efficiency is a challenging problem.

[0003] Therefore, based on the existing work, it is necessary to further explore a method for automatically identifying radar signal modulation modes based on a deep neural network, which can reduce the complexity of the network while ensuring high recognition accuracy, and provide technical support for later transplantation to mobile platforms. Summary of the Invention

[0004] The main technical problem to be solved by the present invention is to provide a method for recognizing radar signal modulation modes with high computational efficiency and accurate signal modulation mode recognition.

[0005] To achieve the above object, the present invention constructs a spatio-temporal feature sharing reconstruction network STARNet to achieve excellent recognition accuracy in a lightweight manner. The present invention provides a method for efficiently recognizing radar signal modulation modes, including the following steps:

[0006] Preprocess the radar I / Q signal to convert it into a radar A / P signal, and perform rotation and occlusion;

[0007] Use the rotated and occluded radar A / P signal as the input of two parallel feature extractors of the efficient spatio-temporal feature sharing reconstruction network to obtain the time-domain feature and spatial-domain feature of the radar A / P signal respectively;

[0008] Fuse the time-domain feature and spatial-domain feature of the radar A / P signal to obtain the spatio-temporal feature of the fused radar signal;

[0009] Use the spatio-temporal feature of the fused radar signal to perform reconstruction and modulation mode recognition tasks through a spatio-temporal sharing decoder module to obtain the final classification result.

[0010] Preferably, the step of preprocessing the radar I / Q signal to convert it into a radar A / P signal and performing rotation and occlusion includes:

[0011] Process the radar I / Q signal into a radar A / P signal, and the processing method is carried out according to the following expression:

[0012]

[0013]

[0014] where r I (n) and r Q (n) are the real and imaginary parts of the I / Q signal, and v A (n) represents the amplitude part of the radar A / P signal, and v P (n) represents the phase part of the radar A / P signal.

[0015] After that, perform a rotation operation on the radar A / P signal, and the processing process is carried out according to the following expression:

[0016]

[0017] where v′ A (n) and v′ P (n) represent the amplitude and phase signals after the rotation operation processing, θ represents the rotation angle, and are respectively set to 0, π,

[0018] Then perform zeroing and random occlusion on the radar A / P signal after the rotation operation processing.

[0019] Preferably, the two parallel feature extractors include: a hybrid attention Ghost extractor (HA-GhostExtractor) and a GRU-based extractor (GRU-based Extractor).

[0020] Preferably, the hybrid attention Ghost extractor (HA-Ghost Extractor) includes: 1 convolutional layer, 3 HA-Ghost modules and 1 global average pooling layer to extract the spatial information of the radar A / P signal and obtain the spatial domain features of the radar A / P signal.

[0021] Among them, the HA-Ghost module improves the conventional Ghost module, changes the kernel size to 1×3 to adapt to the signal sequence, then deploys BN and ReLU to prevent gradient disappearance and improve the training speed, and adds spatial attention on the basis of the original channel attention, which can better suppress irrelevant information in the process of processing radar signals.

[0022] Preferably, the GRU-based extractor (GRU-based Extractor) includes: using 1 two-layer GRU to learn the time correlation of the input signal and obtain the time domain features of the radar A / P signal.

[0023] Preferably, the fusion of the time-domain features and the space-domain features of the radar A / P signal to obtain the spatio-temporal features of the fused radar signal includes:

[0024] The space-domain features extracted by the HA-Ghost extractor and the time-domain features extracted by the GRU-based extractor are fused according to the following expression:

[0025]

[0026] where represents the concatenation operation, F s and F t represent the space-domain features and the time-domain features respectively, and a and b represent the optimal values of the weights, which are used to enhance the useful information and reduce the interference of unimportant information.

[0027] Preferably, the spatio-temporal shared decoder module includes: 1 spatio-temporal shared decoder and 1 modulation classifier. The spatio-temporal shared decoder consists of only 1 convolutional layer and 1 flatten layer, and is used to reconstruct the original input radar A / P signal, which is much smaller than the traditional decoder model. The modulation classifier includes 3 fully connected layers and a softmax layer, and can effectively connect to the fused features and identify the modulation mode.

[0028] Preferably, after the step of reconstructing and identifying the modulation mode of the spatio-temporal features of the fused radar signal through the spatio-temporal shared decoder module to obtain the final classification result, it further includes:

[0029] Using the cross-entropy and the reconstructed mean square error as the total loss function to compare the prediction result with the actual modulation mode label;

[0030] Training and optimizing the modulation classifier by minimizing the total loss function;

[0031] Performing the radar signal modulation mode identification and classification task using the optimized modulation classifier.

[0032] Preferably, the formula of the total loss function is as follows:

[0033]

[0034]

[0035] L f =(1 - α)L clf +αL recon

[0036] where L recon and L clf represent the reconstruction loss and the cross-entropy respectively, Lf represents the total loss. n represents the total number of samples, and v k represents the k-th sample, represents the k-th reconstructed sample. p m equals 1 when belonging to the m-th class, otherwise 0, p m represents the predicted probability of the m-th class. α represents the hyperparameter that controls the balance between the two loss functions.

[0037] In addition, the present invention also provides an efficient radar signal modulation mode recognition device for implementing the above method, including the following modules:

[0038] A signal processing module, configured to preprocess the radar I / Q signal to convert it into a radar A / P signal, and perform rotation and occlusion;

[0039] A feature extraction module, configured to use the rotated and occluded radar A / P signal as the input of two parallel feature extractors of an efficient spatio-temporal feature sharing reconstruction network, and respectively obtain the time domain feature and the spatial domain feature of the radar A / P signal;

[0040] A feature fusion module, configured to fuse the time domain feature and the spatial domain feature of the radar A / P signal to obtain the spatio-temporal feature of the fused radar signal;

[0041] A modulation recognition module, configured to perform reconstruction and modulation mode recognition tasks on the spatio-temporal feature of the fused radar signal through a spatio-temporal shared decoder module to obtain the final classification result.

[0042] The technical solution provided by the present invention has the following beneficial effects:

[0043] (1) The present invention proposes a spatio-temporal feature sharing reconstruction network STARNet. Since this network includes two parallel feature extractors, which respectively use a single autoencoder structure to simultaneously extract the low-dimensional space and time features of the radar signal, it achieves excellent recognition accuracy in a lightweight manner.

[0044] (2) The present invention proposes a hybrid attention module Ghost (HA-Ghost). Since the kernel size is changed to 1×3 to adapt to the signal sequence, and then BN and ReLU are deployed to prevent gradient disappearance and improve the training speed. Based on the original channel attention, spatial attention is added, which can better suppress irrelevant information in the process of processing radar signals. Therefore, it can automatically select more discriminative radar signal information. Compared with the traditional Ghost model, HA-Ghost is more suitable for the automatic modulation mode recognition and classification task.

[0045] (3) A large number of experiments on the benchmark dataset show that the proposed STARNet achieves an average modulation classification accuracy of 63.60%, outperforming the previously state-of-the-art models. Despite extracting more types of features, STARNet has only 14,860 parameters, which is smaller than existing autoencoder-based methods, making it a promising candidate for resource-constrained systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The present invention will be further described below in conjunction with the drawings and embodiments. In the drawings:

[0047] Figure 1 is a flowchart of an efficient radar signal modulation mode recognition method in an embodiment of the present invention;

[0048] Figure 2 is an overall framework diagram of the system in an embodiment of the present invention;

[0049] Figure 3 is a flowchart of preprocessing the radar I / Q signal into a radar A / P signal and performing rotation and occlusion in an embodiment of the present invention;

[0050] Figure 4 is a network structure diagram of STARNet in an embodiment of the present invention;

[0051] Figure 5 is a Ghost module diagram in an embodiment of the present invention, where Figure 5 (a) is a traditional Ghost bottleneck, Figure 5 (b) is an improved HA-Ghost module;

[0052] Figure 6 is a GRU extractor structure diagram in an embodiment of the present invention;

[0053] Figure 7 is a comparison diagram of the overall recognition accuracy of STARNet and other five networks in an embodiment of the present invention;

[0054] Figure 8 is a structure diagram of an efficient radar signal modulation mode recognition device in an embodiment of the present invention;

[0055] Figure 9 is a schematic diagram of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] In order to have a clearer understanding of the technical features, objectives, and effects of the present invention, the specific embodiments of the present invention will now be described in detail with reference to the drawings.

[0057] As Figure 1 、 Figure 2As shown in the figure, an embodiment of the present invention provides an efficient method for identifying radar signal modulation methods, including the following steps:

[0058] S1: Preprocess the radar I / Q signal to convert it into a radar A / P signal, and perform rotation and occlusion.

[0059] It should be understood that the radar I / Q signal refers to two signals received by the radar receiver, representing the real part and the imaginary part of the complex signal respectively, while the radar A / P signal is the information of the radar signal amplitude and phase obtained through formula operations on the existing radar I / Q signal;

[0060] Such as Figure 3 As shown in the figure, first process the original radar I / Q signal into a radar A / P signal, and the processing method is carried out according to the following expression:

[0061]

[0062]

[0063] Where r I (n) and r Q (n) are the real and imaginary parts of the I / Q signal, and v A (n) represents the amplitude part of the radar A / P signal, and v P (n) represents the phase part of the radar A / P signal.

[0064] After that, perform a rotation operation on the radar A / P signal, and the processing process of the rotation operation is carried out according to the following expression:

[0065]

[0066] Where v′ A (n) and v′ P (n) represent the amplitude and phase signals after the rotation operation, and θ represents the rotation angle, which are set to 0, π,

[0067] Then perform zero setting and random occlusion on the radar A / P signal after the rotation operation.

[0068] S2: Propose an efficient spatio-temporal feature sharing reconstruction network (STARNet), and the network structure of the efficient spatio-temporal feature sharing reconstruction network is as Figure 4 As shown in the figure, it includes: two parallel feature extractors (HA-Ghost extractor and GRU-based extractor), a spatio-temporal feature fusion module, and a spatio-temporal sharing decoder module.

[0069] S3: Use the radar A / P signals after rotation and occlusion processing as the inputs of two parallel feature extractors of the efficient spatio-temporal feature sharing reconstruction network, and obtain the time-domain features and spatial-domain features of the radar A / P signals respectively.

[0070] As Figure 5 (a) shows, the traditional Ghost bottleneck usually consists of two Ghost modules, including a depthwise convolution (dconv) layer, a ReLU layer, and a fully connected operation. The dconv layer expands the number of feature map channels through simple linear operations, significantly reducing the computational cost compared with the original convolution. The SE layer connects the two modules to correct the features obtained by the previous module, retaining valuable features and removing worthless features. The 1x1 convolution layer reduces or matches the number of channels, significantly reducing the parameters of the entire module. Specifically, Cin represents the number of input channels, and Cmid and Cout represent the number of output channels generated by the first and second ghost image modules respectively.

[0071] As Figure 5 (b) shows, the present invention further designs a hybrid attention Ghost block with implicit feature reuse based on the Ghost module, making it more suitable for automatic modulation recognition tasks than previous blocks. First, change the kernel size to 1x3 to adapt to the signal sequence, and then deploy BN and ReLU layers to prevent the problem of gradient disappearance and improve the training speed. More importantly, the proposed hybrid attention (HA) module automatically pays attention to distinguish signal information and suppress the influence of noise on the modulated signal sequence under low signal-to-noise ratio conditions.

[0072] Figure 6 The GRU extractor shown uses a two-layer GRU to learn the temporal correlation of the input signal and obtain the time-domain features of the radar A / P signal.

[0073] S4: Fuse the time-domain features and spatial-domain features of the radar A / P signal to obtain the spatio-temporal features of the fused radar signal.

[0074] Fuse the spatial-domain features extracted by the HA-Ghost extractor and the time-domain features extracted by the GRU-based extractor according to the following expression:

[0075]

[0076] Where represents the concatenation operation, F s and F t represent the spatial-domain features and time-domain features respectively, and a and b represent the optimal values of the weights, which are used to enhance useful information and reduce the interference of unimportant information. In this embodiment, a is preferably 0.3 and b is 0.7.

[0077] S5: Reconstruct the spatio-temporal features of the fused radar signal and perform the modulation mode recognition task through the spatio-temporal shared decoder module to obtain the final classification result.

[0078] The spatio-temporal shared decoder module consists of one spatio-temporal shared decoder and one modulation classifier. The spatio-temporal shared decoder is only composed of one convolutional layer and one flatten layer, which is used to reconstruct the original input A / P signal and is much smaller than the traditional decoder model. The modulation classifier includes three fully connected layers and a softmax layer, which can effectively connect to the fused features and identify the modulation mode.

[0079] Further, after the step of reconstructing the spatio-temporal features of the fused radar signal and performing the modulation mode recognition task through the spatio-temporal shared decoder module to obtain the final classification result, it further includes:

[0080] Use cross-entropy and the mean square error of reconstruction as the total loss function to compare the prediction result with the actual modulation mode label;

[0081] Train and optimize the modulation classifier by minimizing the total loss function;

[0082] Use the optimized modulation classifier to perform the radar signal modulation mode recognition and classification task.

[0083] The formula of the total loss function is as follows:

[0084]

[0085]

[0086] L f =(1 - α)L clf +αL recon

[0087] where L recon and L clf represent the reconstruction loss and cross-entropy respectively, L f represents the total loss, n represents the total number of samples, v k represents the kth sample, represents the kth reconstructed sample, p m is equal to 1 when belonging to the mth class, otherwise 0, p m represents the predicted probability of the mth class, and α (taking 0.7) represents the hyperparameter that controls the balance of the two loss functions.

[0088] In this embodiment, the modulation modes of radar signals with 11 modulation formats including BPSK (Binary Phase Shift Keying), 8PSK (8 Phase Shift Keying), QPSK (Quadrature Phase Shift Keying), QAM16 (16 Quadrature Amplitude Modulation), QAM64 (64 Quadrature Amplitude Modulation), BFSK (Binary Frequency Shift Keying), CPFSK (Continuous-phase frequency-shift keying), PAM (Pulse Amplitude Modulation), WB-FM (Wide Band Frequency Modulation), AM-SSB (Amplitude Modulation Single-Side Band), and AM-DSB (Amplitude Modulation Dual-Side Band) were identified. The dataset was Radioml 2016.10a, and the parameters are shown in Table 1:

[0089] Table 1 Simulation Signal Parameters

[0090]

[0091] As Figure 7 shown, Figure 7 This is a comparison chart of the recognition accuracy of the present invention and five other networks under different signal-to-noise ratios. The five networks are respectively:

[0092] DAE (Denoising Auto-Encoder), MCLDNN (Multi-Channel Convolutional Long short-term Deep Neural Network), PET-CGDNN (Parameter Estimation and Transformation based CNN-GRU Deep Neural Network), CLDNN (Convolutional Long short-term Deep Neural Networks), MCNet (A cost-efficient convolutional neural network for robust automatic modulation classification). Tables 2 and 3 show the specific indicators of the recognition rates of the six networks under different signal-to-noise ratio conditions.

[0093] Table 2 Specific indicators of the recognition rates of the six networks under the signal-to-noise ratio conditions of -20 to -2 dB

[0094]

[0095] Table 3 Specific indicators of the recognition rates of the six networks under the signal-to-noise ratio conditions of 0 to 18 dB

[0096]

[0097] From Figure 7 Tables 2 and 3, it can be seen that the recognition accuracy of the STARNet proposed by the present invention is better than that of the other five networks, especially the accuracy at high signal-to-noise ratios.

[0098] Table 4 Specific indicators of the recognition accuracies of the three networks under different signal-to-noise ratios

[0099] Network Number of parameters Floating-point operations per second MCLDNN 406199 665738 DAE 14989 30956 PET 71478 169304 MCNET 90763 89927 CLDNN 248817 588974 STARNet 14860 33886

[0100] To evaluate the computational complexity, the number of learning parameters and the number of floating-point operations per second are summarized in Table 4. It can be seen from Table 4 that the model of the present invention requires the least number of training parameters, which allows it to be used on low-resource terminal devices.

[0101] The present invention discloses an efficient method for identifying radar signal modulation modes, including: preprocessing radar I / Q signals to convert them into radar A / P signals, and performing rotation and occlusion; parallel feature extractors extract time-domain and space-domain features; the fused features are input into a spatio-temporal shared decoder module; the spatio-temporal shared decoder module performs reconstruction and modulation classification tasks. The present invention specifically proposes a spatio-temporal feature sharing reconstruction network STARNet, which achieves excellent recognition accuracy in a lightweight manner. A single autoencoder structure is used to simultaneously extract the low-dimensional spatial and temporal features of radio signals. In addition, a hybrid attention module (HA-Ghost) is designed to automatically extract discriminative radar signal spatial information according to the signal reconstruction performance, thereby reducing the number of model parameters and improving the performance of automatic recognition. A large number of experiments on benchmark datasets show that the proposed STARNet achieves an average modulation classification accuracy of 63.60%, which is better than the previous state-of-the-art models. Moreover, although more types of features are extracted, STARNet has only 14,860 parameters, which is much less than the parameters of existing autoencoder-based methods.

[0102] The following describes an apparatus for identifying an efficient radar signal modulation mode provided by the present invention. The apparatus for identifying an efficient radar signal modulation mode described below can be mutually corresponding and referred to the method for identifying an efficient radar signal modulation mode described above.

[0103] As Figure 8 shown, an apparatus for identifying an efficient radar signal modulation mode includes the following modules:

[0104] A signal processing module 001, configured to preprocess radar I / Q signals to convert them into radar A / P signals, and perform rotation and occlusion;

[0105] A feature extraction module 002, configured to use the radar A / P signals after rotation and occlusion processing as the input of two parallel feature extractors of an efficient spatio-temporal feature sharing reconstruction network, and respectively obtain the time-domain features and space-domain features of the radar A / P signals;

[0106] A feature fusion module 003, configured to fuse the time-domain features and space-domain features of the radar A / P signals to obtain the spatio-temporal features of the fused radar signals;

[0107] A modulation recognition module 004, configured to perform reconstruction and modulation mode recognition tasks on the spatio-temporal features of the fused radar signals through a spatio-temporal shared decoder module to obtain a final classification result.

[0108] Based on but not limited to the above embodiments, the signal processing module 001 is specifically configured to perform the following steps:

[0109] Process the radar I / Q signal into a radar A / P signal, and the processing method is carried out according to the following expression:

[0110]

[0111]

[0112] where r I (n) and r Q (n) are the real and imaginary parts of the radar I / Q signal, and v A (n) represents the amplitude part of the radar A / P signal, and v P (n) represents the phase part of the radar A / P signal;

[0113] Perform a rotation operation on the radar A / P signal, and the rotation operation process is carried out according to the following expression:

[0114]

[0115] where v′ A (n) and v′ P (n) represent the amplitude and phase signals after the rotation operation; θ represents the rotation angle, which is set to 0, π,

[0116] Perform zeroing random occlusion on the rotated radar A / P signal.

[0117] The feature extraction module 002 specifically includes two parallel feature extractors, namely a hybrid attention Ghost extractor and a GRU-based extractor.

[0118] Among them, the hybrid attention Ghost extractor includes: 1 convolutional layer, 3 HA-Ghost modules and 1 global average pooling layer, which are used to extract the spatial information of the radar A / P signal and obtain the spatial domain features of the radar A / P signal;

[0119] Among them, the HA-Ghost module improves the conventional Ghost module, changes the kernel size to 1×3 to adapt to the signal sequence, then deploys BN and ReLU, and adds spatial attention on the basis of the original channel attention.

[0120] Among them, the GRU-based extractor includes: 1 two-layer GRU, which is used to learn the temporal correlation of the input signal and obtain the temporal domain features of the radar A / P signal.

[0121] Based on but not limited to the above embodiments, the feature fusion module 003 is specifically used to perform the following steps:

[0122] The spatial domain features extracted by the hybrid attention Ghost extractor and the temporal domain features extracted by the GRU-based extractor are fused according to the following expression:

[0123]

[0124] where represents the concatenation operation, F s and F t represent the spatial domain features and the temporal domain features respectively, and a and b represent the optimal values of the weights.

[0125] Based on but not limited to the above embodiments, the modulation recognition module 004 is implemented based on the spatio-temporal shared decoder module. The spatio-temporal shared decoder module includes: 1 spatio-temporal shared decoder and 1 modulation classifier;

[0126] The spatio-temporal shared decoder is composed of 1 convolutional layer and 1 flatten layer, and is used to reconstruct the original input radar A / P signal;

[0127] The modulation classifier includes 3 fully connected layers and a softmax layer, and is used to connect to the fused features and identify the modulation mode.

[0128] Based on but not limited to the above embodiments, the modulation recognition module 004 is also used to perform the following steps:

[0129] Compare the prediction result with the actual modulation mode label by using the cross entropy and the reconstructed mean square error as the total loss function;

[0130] Train and optimize the modulation classifier by minimizing the total loss function;

[0131] Use the optimized modulation classifier to perform the radar signal modulation mode recognition and classification task.

[0132] Among them, the formula of the total loss function is as follows:

[0133]

[0134]

[0135] L f =(1 - α)L clf +αL recon

[0136] where L recon and L clf represent the reconstruction loss and the cross entropy respectively, L f represents the total loss, n represents the total number of samples, v k represents the kth sample, represents the kth reconstructed sample, pm is equal to 1 when belonging to the m-th class, and 0 otherwise, p m represents the predicted probability of the m-th class, and α represents the hyperparameter that controls the balance between the two loss functions.

[0137] As Figure 9 shown, a schematic diagram of the physical structure of an electronic device is exemplified. The electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communication bus 640. Among them, the processor 610, the communications interface 620, and the memory 630 complete mutual communication through the communication bus 640. The processor 610 can call the logical instructions in the memory 630 to execute the steps of the above-mentioned efficient radar signal modulation method recognition method, specifically including: preprocessing the radar I / Q signal to convert it into a radar A / P signal, and performing rotation and occlusion; using the rotated and occluded radar A / P signal as the input of two parallel feature extractors of the efficient spatio-temporal feature sharing reconstruction network to respectively obtain the time-domain feature and the spatial-domain feature of the radar A / P signal; fusing the time-domain feature and the spatial-domain feature of the radar A / P signal to obtain the spatio-temporal feature of the fused radar signal; passing the spatio-temporal feature of the fused radar signal through the spatio-temporal sharing decoder module to perform the reconstruction and modulation method recognition task to obtain the final classification result.

[0138] In addition, when the logical instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random 15 Access Memory), a magnetic disk, or an optical disc that can store program codes.

[0139] On the other hand, an embodiment of the present invention also provides a storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned efficient radar signal modulation method are implemented, specifically including: preprocessing the radar I / Q signal to convert it into a radar A / P signal, and performing rotation and occlusion; using the rotated and occluded radar A / P signal as the input of two parallel feature extractors of an efficient spatio-temporal feature sharing reconstruction network to obtain the time-domain feature and the spatial-domain feature of the radar A / P signal respectively; fusing the time-domain feature and the spatial-domain feature of the radar A / P signal to obtain the spatio-temporal feature of the fused radar signal; and performing reconstruction and modulation mode recognition tasks on the spatio-temporal feature of the fused radar signal through a spatio-temporal sharing decoder module to obtain a final classification result.

[0140] It should be noted that in this article, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such a process, method, article or system. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, article or system including that element.

[0141] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. Among the several unit claims listing a number of devices, several of these devices may be embodied by the same hardware item. The use of the terms first, second, and third, etc. does not indicate any order and these terms may be interpreted as identifiers.

[0142] The above is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. An efficient method for identifying radar signal modulation modes, characterized in that, it includes the following steps: Preprocess the radar I / Q signal to convert it into a radar A / P signal, and perform rotation and occlusion; Use the rotated and occluded radar A / P signal as the input of two parallel feature extractors of a spatio-temporal feature sharing reconstruction network to obtain the time domain feature and the spatial domain feature of the radar A / P signal respectively; The two parallel feature extractors include: a hybrid attention Ghost extractor and a GRU-based extractor; The hybrid attention Ghost extractor includes: 1 convolutional layer, 3 HA-Ghost modules and 1 global average pooling layer, which are used to extract the spatial information of the radar A / P signal to obtain the spatial domain feature of the radar A / P signal; Among them, the HA-Ghost module improves the conventional Ghost module, changes the kernel size to 1×3 to adapt to the signal sequence, then deploys BN and ReLU, and adds spatial attention on the basis of the original channel attention; Fuse the time domain feature and the spatial domain feature of the radar A / P signal to obtain the spatio-temporal feature of the fused radar signal; Perform reconstruction and modulation mode recognition tasks on the spatio-temporal feature of the fused radar signal through a spatio-temporal sharing decoder module to obtain the final classification result; The spatio-temporal sharing decoder module includes: 1 spatio-temporal sharing decoder and 1 modulation classifier; Among them, the spatio-temporal sharing decoder consists of 1 convolutional layer and 1 flatten layer, which are used to reconstruct the original input radar A / P signal; The modulation classifier includes 3 fully connected layers and a softmax layer, which are used to connect to the fused features and identify the modulation mode.

2. The efficient method for identifying radar signal modulation modes according to claim 1, characterized in that, The step of preprocessing the radar I / Q signal to convert it into a radar A / P signal and performing rotation and occlusion includes: Process the radar I / Q signal into a radar A / P signal, and the processing method is carried out according to the following expression: where r I (n) and r Q (n) are the real and imaginary parts of the radar I / Q signal, and v A (n) represents the amplitude part of the radar A / P signal, and v P (n) represents the phase part of the radar A / P signal; Perform a rotation operation on the radar A / P signal, and the rotation operation process is carried out according to the following expression: where v' A (n) and v' P (n) represent the amplitude and phase signals after the rotation operation; θ represents the rotation angle, which are respectively set to 0, π, Perform zeroing random occlusion on the rotated radar A / P signal.

3. The efficient method for identifying radar signal modulation modes according to claim 1, characterized in that, The GRU-based extractor includes: 1 two-layer GRU, which is used to learn the time correlation of the input signal to obtain the time domain feature of the radar A / P signal.

4. The efficient method for identifying radar signal modulation modes according to claim 1, characterized in that, The step of fusing the time domain feature and the spatial domain feature of the radar A / P signal to obtain the spatio-temporal feature of the fused radar signal includes: Fuse the spatial domain feature extracted by the hybrid attention Ghost extractor and the time domain feature extracted by the GRU-based extractor according to the following expression: Among them represents a connection operation, F s and F t represent spatial domain features and temporal domain features respectively, and a and b represent the optimal values of the weights.

5. The efficient method for identifying radar signal modulation modes according to claim 1, characterized in that, After the step of performing reconstruction and modulation mode recognition tasks on the spatio-temporal feature of the fused radar signal through a spatio-temporal sharing decoder module to obtain the final classification result, it further includes: The predicted result is compared with the actual modulation mode label by using cross-entropy and the mean square error of reconstruction as the total loss function; The modulation classifier is trained and optimized by minimizing the total loss function; The optimized modulation classifier is used to perform the radar signal modulation mode identification and classification task.

6. The efficient radar signal modulation mode identification method according to claim 5, wherein, the formula of the total loss function is as follows: L f = (1 - α)L clf + αL recon Where L recon and L clf represent the reconstruction loss and the cross entropy respectively, and L f represents the total loss, n represents the total number of samples, v k represents the k-th sample, represents the k-th reconstructed sample, p m is equal to 1 when belonging to the m-th class, otherwise 0, represents the predicted probability of the m-th class, and α represents the hyperparameter that controls the balance between the two loss functions.

7. An efficient radar signal modulation mode identification device for implementing the method according to any one of claims 1-6, wherein, it includes the following modules: A signal processing module, configured to preprocess the radar I / Q signal to convert it into a radar A / P signal, and perform rotation and occlusion; A feature extraction module, configured to use the rotated and occluded radar A / P signal as the input of two parallel feature extractors of an efficient spatio-temporal feature sharing reconstruction network, and respectively obtain the time domain feature and the space domain feature of the radar A / P signal; A feature fusion module, configured to fuse the time domain feature and the space domain feature of the radar A / P signal to obtain the spatio-temporal feature of the fused radar signal; A modulation identification module, configured to perform reconstruction and modulation mode identification tasks on the spatio-temporal feature of the fused radar signal through a spatio-temporal sharing decoder module to obtain a final classification result.

Images