Radar signal modulation format recognition method and device based on time-frequency analysis and denoising
Through the time-frequency analysis method CTFD and the improved ADNet network, combined with the Googlenet network, the problem of radar signal modulation format recognition under low signal-to-noise ratio is solved, high-precision signal denoising and classification recognition are achieved, and the recognition accuracy is significantly improved.
Patent Information
- Application Number
- CN202111579588.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-22
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2041-12-22
AI Technical Summary
In a low signal-to-noise ratio environment, it is difficult for the prior art to accurately identify the modulation format of diverse radar signals, and the traditional denoising method has limited effect on noise interference, which affects the recognition accuracy.
The time-frequency analysis method of Cohen-like time-frequency distribution is used to extract signal characteristics, and the ADNet network is improved for denoising, combined with the improved Googlenet network for signal classification and recognition, and the new kernel function and asymmetric convolution kernel are used to improve the recognition performance.
The recognition accuracy of radar signal modulation format is significantly improved under the low signal-to-noise ratio of -6dB. The improved ADNet network can effectively restore the time-frequency images polluted by noise, and the recognition accuracy can reach 94.81%, which is better than traditional methods.
Smart Images

Figure CN114282576B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of pattern recognition and information technology, and more specifically, to a method and device for identifying a radar signal modulation format based on time-frequency analysis and denoising. Background Art
[0002] At present, advanced electronic reconnaissance technology is the key to gaining an advantage in electronic countermeasures. Radar signal pulse intra-modulation recognition is a key technology for radar electronic warfare (EW). It plays an important role in modern electronic support measures (ESM) systems, electronic intelligence (ELINT) systems, and radar early warning receivers. High-precision recognition of radar signal modulation types means improving the effectiveness of judging the threat level of received signals and improving the accuracy of estimating the parameters of detection signals. However, the pulse compression technology used in radars greatly reduces the power spectral density of radar signals, and the signal-to-noise ratio (SNR) under normal working conditions of radars is also getting lower and lower. This requires that the pulse intra-modulation recognition method of radar signals has good performance under low SNR. In addition, with the rapid development of radar technology, the pulse intra-modulation methods of radar signals are becoming more and more diverse, and the radar signals that need to be identified are more extensive. Therefore, how to accurately identify various radar signal modulation formats in low SNR environments is a hot issue that needs to be solved in the field of radar applications.
[0003] Compared with other signal processing, radar signals have the disadvantage of low signal-to-noise ratio. Therefore, radar signals need to be preprocessed to eliminate the interference of noise signals on the original radar signals. In the process of radar signal denoising, most of them focus on traditional denoising methods, and the use of neural network methods for radar signal denoising and recognition is very limited. In addition, in some studies on neural network denoising of radar signals, the extracted time-frequency features still need to be optimized when the noise is large, and the denoising effect needs to be strengthened. Summary of the invention
[0004] The technical problem to be solved by the present invention is mainly to provide a radar signal modulation recognition method which can achieve obvious denoising effect and accurately identify the signal modulation format.
[0005] In order to achieve the above purpose, the present invention starts from noise resistance, and on the basis of Cohen-like time-frequency distribution, uses the time-frequency analysis method CTFD to extract the time-frequency characteristics of the signal. The new kernel function not only plays a filtering role but also retains the frequency change characteristics of the signal, and can still obtain a time-frequency diagram with high energy concentration in a -6dB low signal-to-noise ratio environment. In order to perform denoising and restoration on the time-frequency diagram polluted by strong noise under low signal-to-noise ratio, the present invention will construct an improved ADNet network for the restoration of noisy time-frequency images. According to Monte Carlo simulation analysis, the time-frequency image polluted by Gaussian or composite noise can be well restored to improve the recognition accuracy. The optimal classification network Googlenet is found and improved in classic classification networks such as Vgg16, Googlenet and Resnet for sorting and recognition of various radar signals.
[0006] According to one aspect of the present invention, the present invention provides a method for identifying the modulation format of a radar signal based on time-frequency analysis and denoising, comprising the following steps:
[0007] Select multiple radar signals;
[0008] Converting each of the radar signals into a time-frequency image by using a time-frequency analysis method CTFD;
[0009] Improve the sparse block of the original ADNet network and set the hole convolution layers in the 2nd, 3rd, 4th and 5th layers of the sparse block;
[0010] De-noising the time-frequency images by using the improved ADNet network to obtain a noise-free time-frequency image;
[0011] Improve the symmetric convolution kernel of the Inception structure in the original Googlenet, and replace the three groups of 7*7 symmetric convolution kernels with three groups of 1*7 and 7*1 asymmetric convolution kernels;
[0012] Signal classification, recognition and prediction are performed on each of the noise-free time-frequency images through the improved Googlenet network to obtain the recognition result of the radar signal modulation format.
[0013] Preferably, the radar signal includes: amplitude modulation SSB-WC, DSB-WC, frequency modulation LFM, SFM, DLFM, MLFM, EQFM, phase modulation BPSK, digital frequency modulation 2FSK, 4FSK, MSK and composite modulation signals LFM-BPSK, 2FSK-BPSK.
[0014] Preferably, the step of converting each radar signal into a time-frequency image by using a time-frequency analysis method CTFD comprises:
[0015] Based on the time-frequency analysis method CWD, the kernel function is improved to obtain the time-frequency analysis method CTFD.
[0016] The expression of the improved kernel function is:
[0017]
[0018] Where τ is the horizontal axis delay, ν is the vertical axis frequency shift, γ is the shape coefficient, and ξ is the scale coefficient, which are used to adjust the shape and size of the kernel function respectively;
[0019] The time-frequency analysis method CTFD is used to perform time-frequency analysis on each of the radar signals to obtain a corresponding time-frequency image.
[0020] Preferably, the step of performing denoising processing on each of the time-frequency images through the improved ADNet network to obtain a noise-free time-frequency image further includes:
[0021] The residual is used to calculate the loss function, and the calculation formula is as follows:
[0022]
[0023] Where N represents the number of time-frequency images, θ is the weight coefficient, is the residual image of the i-th noise-free time-frequency image predicted by the ADNet network, is the noisy time-frequency image of the i-th input, is the ith noise-free time-frequency image, It is the true residual image representing the i-th noise-free time-frequency image. The loss function is calculated by subtracting the true value from the predicted value.
[0024] The improved ADNet network is optimized and trained by minimizing the loss function, and when a preset number of iterations is reached, an optimized ADNet network is obtained.
[0025] Preferably, the step of performing signal classification, recognition and prediction on each of the noise-free time-frequency images through the improved Googlenet network to obtain the recognition result of the radar signal modulation format also includes:
[0026] Information of different scales in the noise-free time-frequency image is extracted by using convolution kernels of different sizes and asymmetric convolution kernels.
[0027] Preferably, after the step of performing signal classification, recognition and prediction on each of the noise-free time-frequency images through the improved Googlenet network to obtain the recognition result of the radar signal modulation format, the method further includes:
[0028] Comparing the recognition result with the real tag to obtain the recognition rate of the radar signal modulation format;
[0029] The recognition performance of the improved Googlenet network is evaluated according to the recognition rate.
[0030] According to another aspect of the present invention, a radar signal modulation format identification device based on time-frequency analysis and denoising is provided, comprising the following modules:
[0031] A selection module is used to select multiple radar signals;
[0032] A conversion module, used for converting each of the radar signals into a time-frequency image by using a time-frequency analysis method CTFD;
[0033] The improvement module is used to improve the sparse block of the original ADNet network and concentrate the hole convolution layers on the 2nd, 3rd, 4th and 5th layers of the sparse block;
[0034] A denoising module, used for denoising each of the time-frequency images through an improved ADNet network to obtain a noise-free time-frequency image;
[0035] The improved module is also used to improve the symmetric convolution kernel of the Inception structure in the original Googlenet, replacing three groups of 7*7 symmetric convolution kernels with three groups of 1*7 and 7*1 asymmetric convolution kernels;
[0036] The recognition module is used to perform signal classification, recognition and prediction on each of the noise-free time-frequency images through the improved Googlenet network to obtain the recognition result of the radar signal modulation format.
[0037] Preferably, the radar signal modulation format identification device based on time-frequency analysis and denoising further includes:
[0038] The evaluation module is used to compare the recognition result with the real label to obtain the recognition rate of the radar signal modulation format, and evaluate the recognition performance of the improved Googlenet network according to the recognition rate.
[0039] The technical solution provided by the present invention has the following beneficial effects:
[0040] (1) Based on the Cohen-like time-frequency distribution, the time-frequency characteristics of the signal are extracted using the time-frequency analysis method CTFD. The new kernel function not only plays a filtering role but also retains the frequency change characteristics of the signal. In the low signal-to-noise ratio environment of -6dB, a time-frequency diagram with high energy concentration can still be obtained.
[0041] (2) In order to restore the time-frequency image polluted by strong noise under low signal-to-noise ratio, the present invention improves the network structure of the ADNet network, and uses the improved ADNet network for the first time to denoise radar signals and restore noisy time-frequency images. Monte Carlo simulation analysis shows that the time-frequency image polluted by Gaussian or composite noise can be well restored, improving the recognition accuracy.
[0042] (3) The optimal classification network Googlenet is improved and used for the sorting and recognition of various radar signals.
[0043] (4) By applying the time-frequency analysis method CTFD, using the improved ADNet network to effectively restore the noisy time-frequency diagram, and using the improved classification network Googlenet for sorting and identification, the recognition accuracy of the radar signal modulation format under low signal-to-noise ratio conditions is significantly improved, and the recognition accuracy can reach 94.81% at -8dB. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:
[0045] Figure 1 It is a flow chart of a radar signal modulation format recognition method based on time-frequency analysis and denoising in an embodiment of the present invention;
[0046] Figure 2 is a kernel function contour map of the time-frequency analysis method CTFD used in the embodiment of the present invention, wherein Figure 2 (a) is the complete kernel function contour map, Figure 2 (b) Figure 2 (a) A partial enlarged view;
[0047] Figure 3 CTFD is a comparison diagram of the time-frequency analysis method for SFM in an embodiment of the present invention and other time-frequency analysis methods, wherein Figure 3 (a) STFT, Figure 3 (b) SET, Figure 3 (c) FSST, Figure 3 (d) MSST (6 iterations), Figure 3 (e) CWD, Figure 3 (f) CTFD;
[0048] Figure 4 The improved ADNet network architecture diagram in the embodiment of the present invention;
[0049] Figure 5 is a structural diagram of an improved sparse block SB in an embodiment of the present invention;
[0050] Figure 6 : is a comparison diagram of the denoising effects of different methods on time-frequency images under -6dB conditions in an embodiment of the present invention, wherein Figure 6 (a) is the image without denoising. Figure 6 (b) is the image denoised using the improved ADNet network. Figure 6 (c) is the image after denoising using singular value decomposition;
[0051] Figure 7 The time-frequency images of 13 radar signals selected in the embodiment of the present invention are denoised by the improved ADNet, where Figure 7 (a) LFM; Figure 7 (b) SFM; Figure 7 (c)DLFM; Figure 7 (d)MLFM; Figure 7 (e) BPSK; Figure 7 (f)SSB-WC; Figure 7 (g)DSB-WC; Figure 7 (h) 2FSK; Figure 7 (i) 4FSK; Figure 7 (j) LFM-BPSK; Figure 7 (k) 2FSK-BPSK; Figure 7 (l)MSK; Figure 7 (m)EQFM;
[0052] Figure 8 This is a schematic diagram of the structure of the improved Inception in the GoogleNet network in an embodiment of the present invention;
[0053] Fig. 9 A comparison chart of recognition accuracies obtained by different time-frequency analysis methods in an embodiment of the present invention;
[0054] Fig.10 This is a comparison chart of recognition accuracy after the signal is not denoised, singular value decomposition is used for denoising, ADNet network is used for denoising, and improved ADNet network is used for denoising in the embodiment of the present invention;
[0055] Fig.11 : This is a comparison chart of recognition accuracy of 13 radar signals before and after denoising in an embodiment of the present invention. Fig.11 (a) LFM; Fig.11 (b) SFM; Fig.11 (c)DLFM; Fig.11 (d)MLFM; Fig.11 (e) BPSK; Fig.11 (f)SSB-WC; Fig.11 (g)DSB-WC; Fig.11(h) 2FSK; Fig.11 (i) 4FSK; Fig.11 (j) LFM-BPSK; Fig.11 (k) 2FSK-BPSK; Fig.11 (l)MSK; Fig.11 (m)EQFM;
[0056] Fig.12 This is a comparison chart of recognition accuracy of different classification networks in an embodiment of the present invention. DETAILED DESCRIPTION
[0057] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.
[0058] This embodiment provides a radar signal modulation format recognition method based on time-frequency analysis and denoising. First, it mainly sorts and identifies a variety of radar modulation signals. In this embodiment, 13 types of radar modulation signals are selected, and the specific number of types can be selected according to actual conditions. Secondly, in order to better classify and identify radar signals under low signal-to-noise ratio conditions, the time-frequency analysis method CTFD is used to perform time-frequency analysis on radar signals to obtain time-frequency images. Third, the improved ADNet network is used to denoise the time-frequency graph of radar signals to obtain clean and noise-free time-frequency graphs. Fourth, optimization is performed in the classification network, and the improved GoogleNet network is used to classify and identify radar signals. Finally, the results of the classification prediction are compared with the actual labels to obtain the recognition rate.
[0059] like Figure 1 As shown, this embodiment provides a radar signal modulation format recognition method based on time-frequency analysis and denoising, which mainly includes the following steps:
[0060] S1. Selection of radar signals;
[0061] High-precision time-frequency analysis and modulation type identification are performed on 13 radar modulation signals, which are: linear frequency modulation signal LFM, sinusoidal frequency modulation signal SFM, dual linear frequency modulation signal DLFM, segmented linear frequency modulation signal MLFM, two-phase coded signal BPSK, single sideband modulation SSB-WC, double sideband modulation DSB-WC, binary frequency shift keying 2FSK, quaternary frequency shift keying 4FSK, linear frequency modulation-two-phase coded composite modulation signal LFM-BPSK, binary frequency shift keying-two-phase coded composite modulation signal 2FSK-BPSK, minimum frequency shift keying MSK, and secondary frequency modulation EQFM. The signal length is N=1024 and the sampling rate is 200MHz. The parameters of each simulation signal are shown in Table 1. The parameters in the table are normalized to the relative sampling rate. For example, the LFM bandwidth is 0.05-0.4, and the actual bandwidth is 10M-80M.
[0062] Table 1 Simulation signal parameters
[0063]
[0064] S2, time-frequency analysis method CTFD extracts time-frequency images;
[0065] The time-frequency analysis method CTFD is a Cohen-type time-frequency analysis method. It is a new method that improves the kernel function based on the time-frequency analysis method CWD. The contour map of the kernel function of the time-frequency analysis method CTFD is shown in the figure below: Figure 2 As shown in the figure, the kernel function is distributed along the υ axis in an elliptical shape, which meets the characteristics of most frequency modulation functions such as sinusoidal frequency modulation signal SFM, quadratic frequency modulation signal EQFM, and phase modulation signal PSK fuzzy domain mutual terms distributed on the υ axis and around it, making the time-frequency analysis method CTFD of the new kernel function have better noise resistance.
[0066] The expression of the improved kernel function is:
[0067]
[0068] Where τ is the horizontal delay, ν is the vertical frequency shift, γ is the shape coefficient, and ξ is the scale coefficient, which is used to adjust the shape and size of the kernel function.
[0069] In order to verify the anti-noise ability of the time-frequency analysis method CTFD, in this embodiment, SFM among 13 radar modulation signals is selected, and the time-frequency analysis method CTFD and other 5 time-frequency analysis methods are used to perform time-frequency analysis on the signal. The time-frequency images obtained are shown in the following figure. Figure 3 As shown, Figure 3 (a) STFT, Figure 3 (b) SET, Figure 3 (c) FSST, Figure 3(d) MSST (6 iterations), Figure 3 (e) CWD, Figure 3 (f) CTFD; By comparison, it can be seen that the time-frequency analysis method CTFD provided by the present invention is significantly better than the other five time-frequency analysis methods.
[0070] S3, improved ADNet network denoising processing;
[0071] like Figure 4 As shown in Figure 1, the improved ADNet network mainly includes: an improved sparse block (SB), a feature enhancement block (FEB), an attention block (AB) and a reconstruction block (RB) for image denoising.
[0072] SB uses common and sparse convolutions to eliminate noise, making a trade-off between performance and efficiency. SB has a total of 12 layers, including two types, dilated Conv+BN+ReLU and Conv+BN+ReLU. Dilated Conv is a hole convolution. Compared with ordinary Conv, dilated Conv increases the receptive field, so that each convolution outputs more information. BN is regularization, ReLU is the activation function, the size of the convolution kernel in the 12 layers is 3*3, the number of input channels is c:3 (color image) or 1 (grayscale image), and the output is 64 channels. The present invention improves the module of SB of the original model, such as Figure 5 As shown in the figure, a better arrangement of dilated convolutions is found. Dilated convolutions are placed in layers 2, 3, 4, and 5 for connection. Compared with the staggered arrangement of the original network, the receptive field of the shallow layer is expanded, so that the network can extract more features in the shallow layer, so that the deep network can have better feature expression. This not only improves the denoising performance and training efficiency of training, but also reduces complexity.
[0073] FEB integrates global and local feature information through a long path to enhance the expressiveness of the denoising model. The 4-layer FEB includes three types: Conv+BN+ReLU, Conv, and Tanh. Conv+BN+ReLU is in the 13th to 15th layers of ADNet, and its convolution kernel size is 64×3×3×64. Conv is located in the 16th layer of ADNet, and its convolution kernel size is 64×3×3×c. Tanh is the activation function. Finally, a Cat operation is used to fuse the original input noisy image with the output of the 16th layer, making the denoising model more expressive.
[0074] AB first uses a 1*1 convolution operation to transform the image of the 18th layer into a picture with c channels, and then multiplies the output of the 16th layer by the obtained weight to extract more prominent noise features.
[0075] RB constructs a clean image through the noise map generated by the network and the given noisy image.
[0076] In the process of denoising of the improved ADNet network, the residual is used to calculate the loss function, and the calculation formula is as follows:
[0077]
[0078] Where N represents the number of time-frequency images, θ is the weight coefficient, is the residual image of the i-th noise-free time-frequency image predicted by the ADNet network, is the noisy time-frequency image of the i-th input, is the ith noise-free time-frequency image, It is the true residual image representing the i-th noise-free time-frequency image. The loss function is calculated by subtracting the true value from the predicted value.
[0079] The improved ADNet network is optimized and trained by minimizing the loss function, and when a preset number of iterations is reached, an optimized ADNet network is obtained.
[0080] In this embodiment, different methods are used to denoise the SFM time-frequency image under the -6dB condition. The denoising effect comparison is shown in the figure below. Figure 6 As shown, Figure 6 : is a comparison diagram of the denoising effects of different methods on time-frequency images under -6dB conditions in an embodiment of the present invention, wherein Figure 6 (a) is the image without denoising. Figure 6 (b) is the image denoised using the improved ADNet network. Figure 6 (c) is the image after denoising using singular value decomposition; by comparison, it can be seen that the time-frequency graph after denoising by the improved ADNet network proposed in the present invention is significantly better than the time-frequency graph after denoising by singular value decomposition.
[0081] In this embodiment, the 13 radar modulation signals are denoised using the improved ADNet network. The denoised time-frequency images corresponding to the 13 modulation modes are shown in FIG. Figure 7 As shown, Figure 7 The time-frequency images of 13 radar signals selected in the embodiment of the present invention are denoised by the improved ADNet, where Figure 7 (a) LFM; Figure 7 (b) SFM; Figure 7 (c)DLFM; Figure 7 (d)MLFM; Figure 7 (e) BPSK; Figure 7 (f)SSB-WC; Figure 7 (g)DSB-WC; Figure 7(h) 2FSK; Figure 7 (i) 4FSK; Figure 7 (j) LFM-BPSK; Figure 7 (k) 2FSK-BPSK; Figure 7 (l)MSK; Figure 7 (m)EQFM; The improved ADNet network proposed in the present invention has good denoising effect on 13 types of radar modulation signals.
[0082] S4, improved GoogleNet for signal classification, recognition and prediction;
[0083] The original GoogleNet itself contains many Inception structures. The present invention improves the Inception structure, such as Figure 8 As shown in the figure, a 1*1 convolution kernel is used to change the corresponding number of channels, and three groups of 7*7 convolution kernels are replaced by three groups of 1*7 and 7*1 asymmetric convolution kernels. This operation can not only reduce the parameters and calculation amount of the network, but also integrate multiple nonlinear activation layers, and improve the discrimination ability to a certain extent. Dropout is used in the network to prevent overfitting. In order to avoid the disappearance of gradients, two auxiliary softmaxes are added to the network to conduct the gradient forward. Compared with other networks, Googlenet enhances the function of the convolution module, which can reduce parameters while increasing the depth and width of the network, reduce the amount of calculation, and ensure the real-time performance of modulation recognition to a certain extent. It is very suitable for classification recognition prediction.
[0084] S5. Compare with the real label to get the recognition rate;
[0085] For each radar signal, when the signal-to-noise ratio is in the range of -10dB to 8dB, 700 samples are simulated as training sets for every 2dB, and 100 samples are simulated as test sets. The training set has a total of 91,000 samples, and the test set has a total of 13,000 samples. The model is trained for a total of 150 epochs, with a training batchsize of 100 and an initial learning rate of 0.01. The learning rate decays to 0.1 times the previous value at epochs 30, 60, and 80. The model reaches the optimal accuracy at the 135th epoch, with the best accuracy of 0.9683.
[0086] like Fig. 9 As shown, Fig. 9 This is a comparison chart of recognition accuracy obtained by different time-frequency analysis methods in the embodiment of the present invention. It can be seen that the recognition rate of the time-frequency analysis method CTFD is improved by more than 30% at a -10dB signal-to-noise ratio, and the recognition rate reaches 100% at a signal-to-noise ratio of -2dB and above.
[0087] like Fig.10 As shown, Fig.10 This is a comparison chart of the recognition accuracy of the embodiments of the present invention when the signal is not denoised, denoised using singular value decomposition, denoised using the ADNet network, and denoised using the improved ADNet network. Although the recognition accuracy of the method proposed in this article is slightly worse than the result of singular value decomposition denoising at a signal-to-noise ratio of -10dB, the recognition accuracy exceeds that of singular value decomposition denoising and ADNet network denoising above -10dB. In addition, the singular value decomposition denoising recognition model only recognizes 8 types of modulation signals, and the present invention expands it to 13 types. At the same time, the BPSK, SSB-WC and DSB-WC signals in the data set have various mixed noises in addition to Gaussian white noise. In this case, the method of the present invention still achieves good recognition accuracy, which fully demonstrates the effectiveness of the method of the present invention and shows its superior performance.
[0088] like Fig.11 As shown, Fig.11 This is a comparison chart of the recognition accuracy of 13 radar signals before and after denoising in the embodiment of the present invention. Except for LFM and MLFM, whose recognition accuracy after denoising is slightly lower than that before denoising, the recognition accuracy of other modulation signals after denoising is, on average, better than that before denoising, especially BPSK, DSB-WC and SSB-WC, which contain mixed noise. When not denoised, the accuracy is very low and the recognition is very difficult. However, after the improved ADNet denoising network, its time-frequency characteristics are greatly enhanced. For example, the accuracy of SSB-WC signal at -10dB signal-to-noise ratio increases from the previous 0.28 to the current 0.75, which is an increase of nearly 50 percentage points. Moreover, the recognition accuracy of the improved ADNet network is better than that of the original ADNet network, which fully demonstrates the effectiveness of the improved ADNet network denoising used in the present invention.
[0089] The present invention also compares the recognition accuracy and recognition speed of the improved Googlenet with the original Googlenet, Vgg16, and Resnet50. The comparison results are shown in Table 2 and Table 3. Fig.12 As shown, Fig.12 This is a comparison chart of the recognition accuracy of different classification networks in the embodiment of the present invention. It can be seen that the improved Googlenet has better recognition speed and accuracy than the original Googlenet under low signal-to-noise ratio conditions, and the recognition accuracy is higher than Resnet50 and Vgg16, and the recognition speed is better than Resnet50 and Vgg16.
[0090] Table 2 Recognition accuracy and recognition speed of different networks
[0091]
[0092] As an optional implementation, a radar signal modulation format identification device based on time-frequency analysis and denoising is provided, comprising the following modules:
[0093] A selection module is used to select multiple radar signals;
[0094] A conversion module, used for converting each of the radar signals into a time-frequency image by using a time-frequency analysis method CTFD;
[0095] The improvement module is used to improve the sparse block of the original ADNet network and concentrate the hole convolution layers on the 2nd, 3rd, 4th and 5th layers of the sparse block;
[0096] A denoising module, used for denoising each of the time-frequency images through an improved ADNet network to obtain a noise-free time-frequency image;
[0097] The improved module is also used to improve the symmetric convolution kernel of the Inception structure in the original Googlenet, replacing three groups of 7*7 symmetric convolution kernels with three groups of 1*7 and 7*1 asymmetric convolution kernels;
[0098] An identification module is used to perform signal classification, identification and prediction on each of the noise-free time-frequency images through an improved Googlenet network to obtain an identification result of the radar signal modulation format;
[0099] The evaluation module is used to compare the recognition result with the real label to obtain the recognition rate of the radar signal modulation format, and evaluate the recognition performance of the improved Googlenet network according to the recognition rate.
[0100] The present invention proposes a radar signal modulation format recognition method and device based on time-frequency analysis and denoising. Starting from the noise resistance, based on the Cohen-like time-frequency distribution, the time-frequency analysis method CTFD is used to extract the time-frequency characteristics of the signal. The new kernel function not only plays a filtering role but also retains the frequency change characteristics of the signal. In the -6dB low signal-to-noise ratio environment, a time-frequency diagram with high energy concentration can still be obtained. In order to perform denoising and restoration of the time-frequency diagram polluted by strong noise under low signal-to-noise ratio, the present invention uses an improved ADNet neural network for the restoration of noisy time-frequency images. Through Monte Carlo simulation analysis, the time-frequency image polluted by Gaussian or composite noise can be well restored to improve the recognition accuracy. The optimal classification network Googlenet is found and improved in the classic classification networks such as Vgg16, Googlenet and Resnet for the sorting and recognition of 13 types of radar signals. By applying a new time-frequency analysis method, using a neural network to effectively restore noisy time-frequency diagrams, and optimizing the classification network, the recognition accuracy of radar signal modulation formats under low signal-to-noise ratio conditions has been significantly improved, and the recognition accuracy can reach 94.81% at -8dB.
[0101] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.
[0102] The serial numbers of the embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In a unit claim that lists several means, several of these means may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order and these words may be interpreted as identifiers.
[0103] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A radar signal modulation format recognition method based on time-frequency analysis and denoising, It is characterized in that The following steps are involved: Select multiple radar signals; Converting each of the radar signals into a time-frequency image by using a time-frequency analysis method CTFD; Improve the sparse block of the original ADNet network and set the hole convolution layers in the 2nd, 3rd, 4th and 5th layers of the sparse block; De-noising the time-frequency images by using the improved ADNet network to obtain a noise-free time-frequency image; Improve the symmetric convolution kernel of the Inception structure in the original Googlenet, using 3 groups and The asymmetric convolution kernel replaces 3 groups Symmetric convolution kernel of ; Using the improved Googlenet network, signal classification, recognition and prediction are performed on each of the noise-free time-frequency images to obtain a recognition result of the radar signal modulation format; The step of converting each of the radar signals into a time-frequency image by using a time-frequency analysis method CTFD comprises: Based on the time-frequency analysis method CWD, the kernel function is improved to obtain the time-frequency analysis method CTFD. The expression of the improved kernel function is: Among them, τ is the horizontal axis delay, ν is the vertical axis frequency shift, γ is the shape coefficient, and ξ is the scale coefficient, which are used to adjust the shape and size of the kernel function respectively; The time-frequency analysis method CTFD is used to perform time-frequency analysis on each of the radar signals to obtain a corresponding time-frequency image.
2. The radar signal modulation format recognition method based on time-frequency analysis and denoising as claimed in claim 1, It is characterized in that The radar signals include: amplitude modulation SSB-WC, DSB-WC, frequency modulation LFM, SFM, DLFM, MLFM, EQFM, phase modulation BPSK, digital frequency modulation 2FSK, 4FSK, MSK and composite modulation signals LFM-BPSK, 2FSK-BPSK.
3. The radar signal modulation format recognition method based on time-frequency analysis and denoising as claimed in claim 1, It is characterized in that The step of performing denoising processing on each of the time-frequency images through the improved ADNet network to obtain a noise-free time-frequency image also includes: The residual is used to calculate the loss function, and the calculation formula is as follows: Where N represents the number of time-frequency images, θ is the weight coefficient, is the residual image of the i-th noise-free time-frequency image predicted by the ADNet network, is the noisy time-frequency image of the i-th input, is the ith noise-free time-frequency image, It is the true residual image representing the i-th noise-free time-frequency image. The loss function is calculated by subtracting the true value from the predicted value. The improved ADNet network is optimized and trained by minimizing the loss function, and when a preset number of iterations is reached, an optimized ADNet network is obtained.
4. The radar signal modulation format recognition method based on time-frequency analysis and denoising as claimed in claim 1, It is characterized in that The step of performing signal classification, recognition and prediction on each of the noise-free time-frequency images through the improved Googlenet network to obtain the recognition result of the radar signal modulation format also includes: Information of different scales in the noise-free time-frequency image is extracted by using convolution kernels of different sizes and asymmetric convolution kernels.
5. The radar signal modulation format recognition method based on time-frequency analysis and denoising as claimed in claim 1, It is characterized in that After the step of performing signal classification, recognition and prediction on each of the noise-free time-frequency images through the improved Googlenet network to obtain the recognition result of the radar signal modulation format, the method further includes: Comparing the recognition result with the real tag to obtain the recognition rate of the radar signal modulation format; The recognition performance of the improved Googlenet network is evaluated according to the recognition rate.
6. A radar signal modulation format recognition device based on time-frequency analysis and denoising, It is characterized in that Includes the following modules: A selection module is used to select multiple radar signals; A conversion module, used for converting each of the radar signals into a time-frequency image by using a time-frequency analysis method CTFD; The improvement module is used to improve the sparse block of the original ADNet network and concentrate the hole convolution layers on the 2nd, 3rd, 4th and 5th layers of the sparse block; A denoising module, used for denoising each of the time-frequency images through an improved ADNet network to obtain a noise-free time-frequency image; The improved module is also used to improve the symmetric convolution kernel of the Inception structure in the original Googlenet, using 3 groups and The asymmetric convolution kernel replaces 3 groups Symmetric convolution kernel of ; An identification module is used to perform signal classification, identification and prediction on each of the noise-free time-frequency images through an improved Googlenet network to obtain an identification result of the radar signal modulation format; The step of converting each of the radar signals into a time-frequency image by using a time-frequency analysis method CTFD comprises: Based on the time-frequency analysis method CWD, the kernel function is improved to obtain the time-frequency analysis method CTFD. The expression of the improved kernel function is: Among them, τ is the horizontal axis delay, ν is the vertical axis frequency shift, γ is the shape coefficient, and ξ is the scale coefficient, which are used to adjust the shape and size of the kernel function respectively; The time-frequency analysis method CTFD is used to perform time-frequency analysis on each of the radar signals to obtain a corresponding time-frequency image.
7. A radar signal modulation format recognition device based on time-frequency analysis and denoising as claimed in claim 6, It is characterized in that Also includes: The evaluation module is used to compare the recognition result with the real label to obtain the recognition rate of the radar signal modulation format, and evaluate the recognition performance of the improved Googlenet network according to the recognition rate.
