CN114282576A - Radar signal modulation format identification method and device based on time-frequency analysis and denoising - Google Patents

Abstract

The invention discloses a radar signal modulation format identification method and a device based on time-frequency analysis and denoising, wherein the method comprises the following steps: selecting radar signals, and converting the signals into time-frequency images by a time-frequency analysis method CTFD; carrying out improved ADNet network denoising treatment; carrying out classification, identification and prediction on signals by the improved Googlenet; and comparing with the real label to obtain the identification rate. The invention starts from noise immunity, utilizes CTFD to extract the time-frequency characteristics of signals, and a new kernel function also reserves the frequency change characteristics of the signals while filtering, and constructs an improved ADNet network for restoring a noise-containing time-frequency image in order to carry out noise reduction and recovery on a time-frequency image polluted by strong noise under a low signal-to-noise ratio. The method can well restore the time-frequency image polluted by Gaussian or composite noise, and improve the identification precision. And constructing an improved Googlenet network for sorting and identifying various radar signals. The identification precision of the radar signal modulation format under the condition of low signal-to-noise ratio is remarkably improved, and the identification accuracy can reach 94.81% under-8 dB.

Description

Radar signal modulation format identification method and device based on time-frequency analysis and denoising

Technical Field

The invention belongs to the technical field of pattern recognition and information, and particularly relates to a radar signal modulation format recognition method and device based on time-frequency analysis and denoising.

Background

Currently, advanced electronic reconnaissance techniques are key to gaining advantages in electronic warfare. The identification of the modulation in the radar signal pulse is a key technology of radar Electronic Warfare (EW). Plays an important role in modern Electronic Support Measures (ESM) systems, electronic intelligence (elint) systems and radar warning receivers. The high-precision identification of the radar signal modulation type means that the effectiveness of judging the threat level of the received signal is improved, and the accuracy of estimating the parameters of the detection signal is improved. However, the power spectral density of the radar signal is greatly reduced by the pulse compression technique used in the radar, and the signal-to-noise ratio (SNR) under the normal operating environment of the radar is also increasingly lower. This requires that the method of identification of the intra-pulse modulation of the radar signal has good performance at low signal-to-noise ratios. In addition, with the rapid development of radar technology, the pulse internal modulation mode of radar signals is more and more diversified, and the radar signals needing to be identified are wider. Therefore, how to accurately identify various different radar signal modulation formats in the low signal-to-noise ratio environment is a hot point problem to be solved urgently in the field of radar application.

Radar signals have the disadvantage of a low signal-to-noise ratio compared to other signal processing. Therefore, the radar signal needs to be preprocessed to eliminate the interference of the noise signal to the original radar signal. In the de-noising process of the radar signals, the traditional de-noising method is mostly focused on, and the work of de-noising and identifying the radar signals by utilizing the neural network method is carried out in a limited way. In addition, in some researches on neural network denoising of radar signals, extracted time-frequency features still need to be optimized when noise is large, and denoising effects need to be enhanced.

Disclosure of Invention

The invention mainly solves the technical problem of providing a radar signal modulation and identification method which has obvious denoising effect and can accurately identify a signal modulation format.

In order to achieve the purpose, the invention uses a time-frequency analysis method CTFD to extract the time-frequency characteristics of the signals on the basis of Cohen time-frequency distribution from the viewpoint of noise immunity, and the new kernel function plays a role in filtering and simultaneously retains the frequency change characteristics of the signals, so that a time-frequency graph with high energy concentration can be obtained under the environment of-6 dB low signal-to-noise ratio. In order to carry out noise reduction and recovery on the time-frequency image polluted by strong noise under the condition of low signal-to-noise ratio, the invention constructs an improved ADNet network for restoring the time-frequency image containing noise. By Monte Carlo simulation analysis, the time-frequency image polluted by Gaussian or composite noise can be well restored, and the identification precision is improved. The optimization is realized in classical classification networks such as Vgg16, Googlenet and Resnet, and an optimal classification network Googlenet is found and improved for sorting and identifying various radar signals.

According to one aspect of the invention, the invention provides a radar signal modulation format identification method based on time-frequency analysis and denoising, which comprises the following steps:

selecting a plurality of radar signals;

converting each radar signal into a time-frequency image by a time-frequency analysis method CTFD;

improving the sparse block of the original ADNet network, and intensively arranging the hole convolution layers on the 2 nd, 3 rd, 4 th and 5 th layers of the sparse block;

denoising each time-frequency image through an improved ADNet network to obtain a noiseless time-frequency image;

modifying the symmetric convolution kernel of the inclusion structure in the original Googlenet, and replacing 3 groups of symmetric convolution kernels of 7 by 3 groups of asymmetric convolution kernels of 1 by 7 and 7 by 1;

and performing signal classification, identification and prediction on each noiseless time-frequency image through an improved Googlenet network to obtain an identification result of a radar signal modulation format.

Preferably, the radar signal comprises: amplitude modulation SSB-WC, DSB-WC, frequency modulation LFM, SFM, DLFM, MLFM, EQFM, phase modulation BPSK, digital frequency modulation 2FSK, 4FSK, MSK and complex modulation signals LFM-BPSK, 2 FSK-BPSK.

Preferably, the step of converting each radar signal into a time-frequency image by a time-frequency analysis method CTFD includes:

improving a kernel function on the basis of a CWD (time-frequency analysis) method to obtain a CTFD (time-frequency analysis);

the expression of the improved kernel function is as follows:

wherein tau is time delay of horizontal axis, nu is frequency shift of vertical axis, gamma is shape coefficient, xi is scale coefficient, used for adjusting shape and size of kernel function separately;

and performing time-frequency analysis on each radar signal by the time-frequency analysis method CTFD to obtain a corresponding time-frequency image.

Preferably, the step of denoising each time-frequency image through the improved ADNet network to obtain a noise-free time-frequency image further includes:

the residual is used to calculate a loss function, which is calculated as follows:

wherein N represents the number of time-frequency images, theta is a weight coefficient,

is a residual error map of the ith noiseless time-frequency image predicted by the ADNet network,

for the ith input noisy time-frequency image,

for the ith noise-free time-frequency image,

the method comprises the steps of calculating a loss function by subtracting a true value from a predicted value, wherein the true residual map represents an ith noiseless time-frequency image;

and carrying out optimization training on the improved ADNet network by minimizing the loss function, and obtaining the optimized ADNet network when preset iteration times are reached.

Preferably, the step of performing signal classification, identification and prediction on each noiseless time-frequency image through an improved Googlenet network to obtain an identification result of a radar signal modulation format further includes:

and extracting information of different scales in the noiseless time-frequency image through convolution kernels with different sizes and asymmetric convolution kernels.

Preferably, after the step of performing signal classification, identification and prediction on each noiseless time-frequency image through the improved Googlenet network to obtain an identification result of a radar signal modulation format, the method further includes:

comparing the identification result with a real label to obtain the identification rate of the radar signal modulation format;

and evaluating the recognition performance of the improved Googlenet network according to the recognition rate.

According to another aspect of the present invention, there is also provided a radar signal modulation format recognition apparatus based on time-frequency analysis and denoising, including the following modules:

the selecting module is used for selecting various radar signals;

the conversion module is used for converting each radar signal into a time-frequency image by a time-frequency analysis method CTFD;

the improvement module is used for improving the sparse block of the original ADNet network and intensively arranging the cavity convolution layer on the 2 nd, 3 rd, 4 th and 5 th layers of the sparse block;

the denoising module is used for denoising each time-frequency image through the improved ADNet network to obtain a noise-free time-frequency image;

the improvement module is also used for improving the symmetric convolution kernel of the inclusion structure in the original Googlenet, and replacing 3 groups of symmetric convolution kernels 7 by 7 with 3 groups of asymmetric convolution kernels 1 by 7 and 7 by 1;

and the identification module is used for carrying out signal classification, identification and prediction on each noiseless time-frequency image through an improved Googlenet network to obtain an identification result of a radar signal modulation format.

Preferably, the radar signal modulation format recognition apparatus based on time-frequency analysis and denoising further includes:

and the evaluation module is used for comparing the identification result with a real label to obtain the identification rate of the radar signal modulation format, and evaluating the identification performance of the improved Googlenet network according to the identification rate.

The technical scheme provided by the invention has the following beneficial effects:

(1) on the basis of Cohen time frequency distribution, a time frequency analysis method CTFD is used for extracting time frequency characteristics of signals, the new kernel function plays a role in filtering and simultaneously retains frequency change characteristics of the signals, and a time frequency graph with high energy concentration can be obtained under a-6 dB low signal-to-noise ratio environment.

(2) In order to carry out noise reduction and recovery on the time-frequency image polluted by strong noise under the condition of low signal-to-noise ratio, the invention improves the network structure of the ADNet network, and uses the improved ADNet network for denoising the radar signal for the first time and restoring the noise-containing time-frequency image. Monte Carlo simulation analysis shows that the time-frequency image polluted by Gaussian or composite noise can be well restored, and the identification precision is improved.

(3) The optimization is realized in classical classification networks such as Vgg16, Googlenet and Resnet, and the optimal classification network Googlenet is improved and used for sorting and identifying various radar signals.

(4) By applying the time-frequency analysis method CTFD, effectively recovering the noisy time-frequency diagram by using the improved ADNet network and carrying out sorting and identification by using the improved classification network Googlenet, the identification precision of the radar signal modulation format under the condition of low signal-to-noise ratio is obviously improved, and the identification accuracy can reach 94.81 percent under-8 dB.

Drawings

The invention will be further described with reference to the accompanying drawings and examples, in which:

FIG. 1 is a flow chart of a radar signal modulation format recognition method based on time-frequency analysis and denoising in the embodiment of the present invention;

FIG. 2 is a contour diagram of kernel function of a time-frequency analysis method CTFD used in an embodiment of the present invention, wherein FIG. 2(a) is a complete contour diagram of kernel function, and FIG. 2(b) is a partial enlarged view of FIG. 2 (a);

fig. 3 is a comparison diagram of a time-frequency analysis method CTFD and other time-frequency analysis methods for SFM according to an embodiment of the present invention, where fig. 3(a) STFT, fig. 3(b) SET, fig. 3(c) FSST, fig. 3(d) MSST (iteration 6), fig. 3(e) CWD, and fig. 3(f) CTFD;

FIG. 4 is a diagram of the architecture of an ADNet network modified in accordance with an embodiment of the present invention;

FIG. 5 is a structural diagram of an improved sparse block SB in an embodiment of the present invention;

FIG. 6 is a comparison graph of denoising effects of time-frequency images by different methods under the-6 dB condition in the embodiment of the present invention, where FIG. 6(a) is an image without denoising, FIG. 6(b) is an image denoised by using an improved ADNet network, and FIG. 6(c) is an image denoised by using singular value decomposition;

fig. 7 is a diagram of the effect of time-frequency images of 13 selected radar signals after ADNet denoising is performed in an improved manner in the embodiment of the present invention, where fig. 7(a) LFM; FIG. 7(b) SFM; FIG. 7(c) DLFM; FIG. 7(d) MLFM; fig. 7(e) BPSK; FIG. 7(f) SSB-WC; FIG. 7(g) DSB-WC; fig. 7(h)2 FSK; FIG. 7(i)4 FSK; FIG. 7(j) LFM-BPSK; FIG. 7(k)2 FSK-BPSK; FIG. 7(l) MSK; FIG. 7(m) EQFM;

fig. 8 is a schematic structural diagram of an improved inclusion in the GoogleNet network in the embodiment of the present invention;

FIG. 9 is a comparison graph of recognition accuracy obtained by different time-frequency analysis methods in an embodiment of the present invention;

FIG. 10 is a comparison graph of recognition accuracy after denoising using an ADNet network and a modified ADNet network without denoising a signal according to an embodiment of the present invention;

FIG. 11 is a comparison graph of recognition accuracy before and after de-noising of 13 radar signals in an embodiment of the present invention, where 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)2 FSK; FIG. 11(i)4 FSK; FIG. 11(j) LFM-BPSK; FIG. 11(k)2 FSK-BPSK; FIG. 11(l) MSK; FIG. 11(m) EQFM;

FIG. 12 is a comparison diagram of recognition accuracy of different classification networks according to an embodiment of the present invention.

Detailed Description

For a more clear understanding of the technical features, objects and effects of the present invention, embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

The embodiment provides a radar signal modulation format identification method based on time-frequency analysis and denoising, and the method is mainly used for sorting and identifying various radar modulation signals. Secondly, in order to better classify and identify the radar signals under the condition of low signal-to-noise ratio, a time-frequency analysis method CTFD is used for carrying out time-frequency analysis on the radar signals to obtain a time-frequency image. Thirdly, the improved ADNet network is used for denoising the radar signal time-frequency graph to obtain a clean noise-free time-frequency graph. Fourthly, the optimization is carried out in a classification network, and the classification and identification of the radar signals are carried out by using the improved GoogleNet network. And finally, comparing the classified and predicted result with a real label to obtain the recognition rate.

As shown in fig. 1, this embodiment provides a radar signal modulation format identification method based on time-frequency analysis and denoising, which mainly includes the following steps:

s1, selecting a radar signal;

high-precision time-frequency analysis and modulation type identification are carried out on 13 radar modulation signals, wherein the 13 radar modulation signals are respectively as follows: linear frequency modulation signal LFM, sinusoidal frequency modulation signal SFM, bilinear frequency modulation signal DLFM, piecewise linear frequency modulation signal MLFM, binary 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-binary coded composite modulated signal LFM-BPSK, binary frequency shift keying-binary coded composite modulated signal 2FSK-BPSK, minimum frequency shift keying MSK, secondary frequency modulation EQFM. The signal length is N1024, and the sampling rate is 200 MHz. The simulation signal parameters are shown in table 1, and the parameters in the table are normalized to the relative sampling rate, for example, the bandwidth of LFM is 0.05-0.4, and the actual bandwidth is 10M-80M.

TABLE 1 simulation Signal parameters

S2, extracting a time-frequency image by a time-frequency analysis method (CTFD);

the time-frequency analysis method CTFD is a Cohen time-frequency analysis method, and is a new method for improving a kernel function on the basis of a time-frequency analysis method CWD, a contour map of the kernel function of the time-frequency analysis method CTFD is shown in figure 2, the kernel function is distributed along a upsilon axis and is elliptic, the characteristics that most frequency modulation functions such as a sinusoidal frequency modulation signal SFM, a secondary frequency modulation signal EQFM and a phase modulation signal PSK fuzzy domain are distributed on the upsilon axis and the periphery of the upsilon axis are met, and the time-frequency analysis method CTFD of the new kernel function has better anti-noise capability.

The expression of the improved kernel function is as follows:

wherein tau is horizontal axis time delay, nu is vertical axis frequency shift, gamma is shape coefficient, xi is scale coefficient, and is used for adjusting the shape and size of the kernel function.

In order to verify the anti-noise capability of the time-frequency analysis method CTFD, in this embodiment, SFM of 13 radar modulation signals is selected, the time-frequency analysis method CTFD and other 5 time-frequency analysis methods are respectively adopted to perform time-frequency analysis on the signals, and the obtained time-frequency images are shown in fig. 3, where fig. 3(a) STFT, fig. 3(b) SET, fig. 3(c) FSST, fig. 3(d) MSST (iteration 6 times), fig. 3(e) CWD, and fig. 3(f) CTFD; the comparison shows that the time-frequency analysis method CTFD provided by the invention is obviously superior to other 5 time-frequency analysis methods.

S3, carrying out improved ADNet network denoising processing;

as shown in fig. 4, the improved ADNet network mainly includes: a modified Sparse Block (SB), a Feature Enhancement Block (FEB), an Attention Block (AB), and a Reconstruction Block (RB) for image denoising.

SB trades off performance against efficiency by using common and sparse convolutions to cancel noise. The SB has 12 layers in total, including two types, scaled Conv + BN + ReLU and Conv + BN + ReLU, the scaled Conv is a hole convolution, compared with the common Conv, the scaled Conv enlarges the sensing field, so that more BN outputted by each convolution is regularized, the ReLU is an activation function, the size of convolution kernels in the 12 layers is 3 x 3, the number of input channels is c:3 (color graph) or 1 (gray graph), and the output is 64 channels. This not only improves the denoising performance and training efficiency of training, but also reduces the complexity.

The FEB integrates global characteristic information and local characteristic information through a long path to enhance the expression capability of a denoising model. The 4-layer FEB contains three types Conv + BN + ReLU, Conv and Tanh, wherein Conv + BN + ReLU are in 13-15 layers of ADNet, and the convolution kernel size is 64 multiplied by 3 multiplied by 64. Conv is at layer 16 in ADNet, with convolution kernel size of 64 × 3 × 3 × c. Tanh is the activation function. And finally, fusing the original input noisy image and the 16 th layer output by using a Cat operation, so that the denoise model has stronger expressive ability.

AB first changes the 18 th layer image into a picture with channel number c using a 1 x 1 convolution operation, and then multiplies the 16 th layer output by the obtained weight to extract more prominent noise features.

RB constructs a clean image through the network generated noise map and the given noise image.

In the process of the improved ADNet network denoising treatment, a residual error is used for calculating a loss function, and the calculation formula is as follows:

wherein N represents the number of time-frequency images, theta is a weight coefficient,

is a residual error map of the ith noiseless time-frequency image predicted by the ADNet network,

for the ith input noisy time-frequency image,

for the ith noise-free time-frequency image,

the method comprises the steps of calculating a loss function by subtracting a true value from a predicted value, wherein the true residual map represents an ith noiseless time-frequency image;

and carrying out optimization training on the improved ADNet network by minimizing the loss function, and obtaining the optimized ADNet network when preset iteration times are reached.

In this embodiment, different methods are respectively adopted to denoise an SFM time-frequency image under a-6 dB condition, a denoising effect contrast graph is shown in fig. 6, fig. 6 is a denoising effect contrast graph of a time-frequency image under a-6 dB condition in the embodiment of the present invention, where fig. 6(a) is an image that is not denoised, fig. 6(b) is an image denoised by adopting an improved ADNet network, and fig. 6(c) is an image denoised by using a singular value decomposition; the comparison shows that the time-frequency graph denoised by the improved ADNet network provided by the invention is obviously superior to the time-frequency graph denoised by singular value decomposition.

In this embodiment, the 13 radar modulation signals are denoised by an improved ADNet network, the denoised time-frequency images corresponding to the 13 modulation modes are shown in fig. 7, fig. 7 is an effect diagram of the time-frequency images of the 13 radar signals selected in the embodiment of the present invention after being denoised by the improved ADNet, wherein fig. 7(a) LFM; FIG. 7(b) SFM; FIG. 7(c) DLFM; FIG. 7(d) MLFM; fig. 7(e) BPSK; FIG. 7(f) SSB-WC; FIG. 7(g) DSB-WC; fig. 7(h)2 FSK; FIG. 7(i)4 FSK; FIG. 7(j) LFM-BPSK; FIG. 7(k)2 FSK-BPSK; FIG. 7(l) MSK; FIG. 7(m) EQFM; the improved ADNet network provided by the invention has a good denoising effect on 13 radar modulation signals.

S4, carrying out signal classification, identification and prediction on the improved GoogleNet;

the original GoogleNet contains a plurality of inclusion structures, the inclusion structures are improved, as shown in figure 8, a convolution kernel of 1 x 1 is used for changing the corresponding number of channels, 3 groups of asymmetric convolution kernels of 1 x 7 and 7 x 1 are used for replacing a convolution kernel of 3 groups of 7 x 7, and by adopting the operation, the parameters and the calculated amount of the network can be reduced, a plurality of nonlinear active layers are integrated, and the discrimination capability is improved to a certain extent. Dropout in the network is used to prevent overfitting, and in order to avoid the gradient vanishing, 2 additional softmax in the network are added for conducting the gradient forward. Compared with other networks, the Googlenet enhances the function of a convolution module, can reduce parameters while increasing the depth and width of the network, reduces the calculated amount, and ensures the real-time property of modulation identification to a certain extent. Is suitable for classification recognition prediction.

S5, comparing the real label to obtain the identification rate;

for each radar signal, when the signal-to-noise ratio is in the range of-10 dB to 8dB, 700 samples are simulated every 2dB to serve as a training set, 100 samples are simulated to serve as a test set, the training set comprises 91000 samples in total, 13000 samples in total are simulated in the test set, the model is trained for 150 epochs in total, the training Batchsize is 100, the initial learning rate is 0.01, the learning rate is attenuated to 0.1 times before when the epochs are 30, 60 and 80, the precision of the model reaches the optimum at the 135 th epoch, and the optimum precision is 0.9683.

As shown in fig. 9, fig. 9 is a comparison diagram 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% in a whole way under the signal-to-noise ratio of-10 dB, and the recognition rate reaches 100% under the signal-to-noise ratio of-2 dB or above.

As shown in fig. 10, fig. 10 is a comparison graph of recognition accuracy after denoising by using ADNet network and denoising by using singular value decomposition without denoising a signal according to the embodiment of the present invention. Although the recognition accuracy of the method is slightly worse than the result of singular value decomposition denoising under the signal-to-noise ratio of-10 dB, the recognition accuracy exceeds the singular value decomposition denoising and ADNet network denoising above-10 dB. In addition, the singular value decomposition denoising identification model only identifies 8 modulation signals, the method expands the number of modulation signals to 13, simultaneously BPSK, SSB-WC and DSB-WC signals in a data set have various mixed noises besides Gaussian white noise, and under the condition, the method still obtains good identification precision, fully demonstrates the effectiveness of the method and expresses the superior performance of the method.

As shown in fig. 11, fig. 11 is a comparison diagram of recognition accuracy before and after de-noising 13 radar signals in the embodiment of the present invention. Except LFM and MLFM denoised identification precision is slightly reduced compared with that of non-denoised signals, on average, other modulation signals are denoised, the identification precision is better than that of non-denoised signals, particularly, BPSK, DSB-WC and SSB-WC which contain mixed noise are low in precision and high in identification difficulty when the signals are not denoised, but after the ADNet denoising network is improved, the time-frequency characteristic is greatly enhanced, for example, the precision of the SSB-WC signal is increased from 0.28 to 0.75 when the signal is subjected to-10 dB signal to noise ratio, and the integral improvement is nearly 50 percentage points. In addition, the recognition precision of the improved ADNet network is superior to that of the original ADNet network, and the effectiveness of the improved ADNet network denoising used by the invention is fully demonstrated.

The invention also compares the recognition accuracy and the recognition speed of the improved Googlenet with the original Googlenet, Vgg16 and Resnet50, and the comparison result is shown in table 2 and fig. 12, where fig. 12 is a comparison graph of the recognition accuracy of different classification networks in the embodiment of the invention. It can be seen that the improved Googlenet has better recognition speed and accuracy than the original Googlenet under the condition of low signal-to-noise ratio, the recognition accuracy is higher than Resnet50 and Vgg16, and the recognition speed is better than Resnet50 and Vgg 16.

TABLE 2 different network recognition accuracy and recognition speed

As an optional implementation manner, a radar signal modulation format recognition apparatus based on time-frequency analysis and denoising is provided, which includes the following modules:

the selecting module is used for selecting various radar signals;

the conversion module is used for converting each radar signal into a time-frequency image by a time-frequency analysis method CTFD;

the improvement module is used for improving the sparse block of the original ADNet network and intensively arranging the cavity convolution layer on the 2 nd, 3 rd, 4 th and 5 th layers of the sparse block;

the denoising module is used for denoising each time-frequency image through the improved ADNet network to obtain a noise-free time-frequency image;

the improvement module is also used for improving the symmetric convolution kernel of the inclusion structure in the original Googlenet, and replacing 3 groups of symmetric convolution kernels 7 by 7 with 3 groups of asymmetric convolution kernels 1 by 7 and 7 by 1;

the identification module is used for carrying out signal classification, identification and prediction on each noiseless time-frequency image through an improved Googlenet network to obtain an identification result of a radar signal modulation format;

and the evaluation module is used for comparing the identification result with a real label to obtain the identification rate of the radar signal modulation format, and evaluating the identification performance of the improved Googlenet network according to the identification rate.

The invention provides a radar signal modulation format identification method and device based on time-frequency analysis and denoising, which are based on noise immunity, on the basis of Cohen time-frequency distribution, a time-frequency analysis method CTFD is used for extracting time-frequency characteristics of signals, a new kernel function plays a role in filtering, frequency change characteristics of the signals are also reserved, and a time-frequency graph with high energy concentration can be obtained under a-6 dB low signal-to-noise ratio environment. In order to carry out noise reduction and recovery on the time-frequency image polluted by strong noise under the condition of low signal-to-noise ratio, the invention utilizes an improved ADNet neural network to restore the time-frequency image containing noise. By Monte Carlo simulation analysis, the time-frequency image polluted by Gaussian or composite noise can be well restored, and the identification precision is improved. The optimization is realized in classical classification networks such as Vgg16, Googlenet and Resnet, and the optimal classification network Googlenet is found and improved for sorting and identifying 13 radar signals. By applying a new time-frequency analysis method, effectively recovering a noisy time-frequency graph by using a neural network and optimizing a classification network, the identification precision of the radar signal modulation format under the condition of low signal-to-noise ratio is remarkably improved, and the identification accuracy can reach 94.81 percent under-8 dB.

It should be noted that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or system. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other like elements in a process, method, article, or system that comprises the element.

The above-mentioned serial numbers of the embodiments of the present invention are merely for description and do not represent the merits of the embodiments. In the unit claims enumerating several means, several of these means may be embodied by one and the same item of hardware. The use of the words first, second, third and the like do not denote any order, but rather the words first, second and the like may be interpreted as indicating any order.

The above description is only a preferred embodiment of the present invention, and not intended to limit the scope of the present invention, and all modifications of equivalent structures and equivalent processes, which are made by using the contents of the present specification and the accompanying drawings, or directly or indirectly applied to other related technical fields, are included in the scope of the present invention.

Claims (8)

1. A radar signal modulation format identification method based on time-frequency analysis and denoising is characterized by comprising the following steps:

selecting a plurality of radar signals;

converting each radar signal into a time-frequency image by a time-frequency analysis method CTFD;

improving the sparse block of the original ADNet network, and intensively arranging the hole convolution layers on the 2 nd, 3 rd, 4 th and 5 th layers of the sparse block;

denoising each time-frequency image through an improved ADNet network to obtain a noiseless time-frequency image;

modifying the symmetric convolution kernel of the inclusion structure in the original Googlenet, and replacing 3 groups of symmetric convolution kernels of 7 by 3 groups of asymmetric convolution kernels of 1 by 7 and 7 by 1;

and performing signal classification, identification and prediction on each noiseless time-frequency image through an improved Googlenet network to obtain an identification result of a radar signal modulation format.

2. The method for recognizing radar signal modulation format based on time-frequency analysis and denoising as claimed in claim 1, wherein the radar signal comprises: amplitude modulation SSB-WC, DSB-WC, frequency modulation LFM, SFM, DLFM, MLFM, EQFM, phase modulation BPSK, digital frequency modulation 2FSK, 4FSK, MSK and complex modulation signals LFM-BPSK, 2 FSK-BPSK.

3. The method for radar signal modulation format recognition based on time-frequency analysis and denoising of claim 1, wherein the step of converting each radar signal into a time-frequency image by a time-frequency analysis method CTFD comprises:

improving a kernel function on the basis of a CWD (time-frequency analysis) method to obtain a CTFD (time-frequency analysis);

the expression of the improved kernel function is as follows:

wherein, tau is time delay of horizontal axis, nu is frequency shift of vertical axis, gamma is shape coefficient, xi is scale coefficient, used for adjusting shape and size of kernel function separately;

and performing time-frequency analysis on each radar signal by the time-frequency analysis method CTFD to obtain a corresponding time-frequency image.

4. The method for identifying a radar signal modulation format based on time-frequency analysis and denoising of claim 1, wherein the step of denoising each of the time-frequency images through an improved ADNet network to obtain a noise-free time-frequency image further comprises:

the residual is used to calculate a loss function, which is calculated as follows:

wherein N represents the number of time-frequency images, theta is a weight coefficient,

is a residual error map of the ith noiseless time-frequency image predicted by the ADNet network,

for the ith input noisy time-frequency image,

for the ith noise-free time-frequency image,

the method comprises the steps of calculating a loss function by subtracting a true value from a predicted value, wherein the true residual map represents an ith noiseless time-frequency image;

and carrying out optimization training on the improved ADNet network by minimizing the loss function, and obtaining the optimized ADNet network when preset iteration times are reached.

5. The radar signal modulation format recognition method based on time-frequency analysis and denoising of claim 1, wherein the step of performing signal classification recognition prediction on each noiseless time-frequency image through an improved Googlenet network to obtain a recognition result of a radar signal modulation format further comprises:

and extracting information of different scales in the noiseless time-frequency image through convolution kernels with different sizes and asymmetric convolution kernels.

6. The method for recognizing radar signal modulation format based on time-frequency analysis and denoising as claimed in claim 1, wherein after the step of performing signal classification recognition prediction on each noiseless time-frequency image through the improved Googlenet network to obtain the recognition result of radar signal modulation format, the method further comprises:

comparing the identification result with a real label to obtain the identification rate of the radar signal modulation format;

and evaluating the recognition performance of the improved Googlenet network according to the recognition rate.

7. A radar signal modulation format recognition device based on time-frequency analysis and denoising is characterized by comprising the following modules:

the selecting module is used for selecting various radar signals;

the conversion module is used for converting each radar signal into a time-frequency image by a time-frequency analysis method CTFD;

the improvement module is used for improving the sparse block of the original ADNet network and intensively arranging the cavity convolution layer on the 2 nd, 3 rd, 4 th and 5 th layers of the sparse block;

the denoising module is used for denoising each time-frequency image through the improved ADNet network to obtain a noise-free time-frequency image;

the improvement module is also used for improving the symmetric convolution kernel of the inclusion structure in the original Googlenet, and replacing 3 groups of symmetric convolution kernels 7 by 7 with 3 groups of asymmetric convolution kernels 1 by 7 and 7 by 1;

and the identification module is used for carrying out signal classification, identification and prediction on each noiseless time-frequency image through an improved Googlenet network to obtain an identification result of a radar signal modulation format.

8. The apparatus for recognizing radar signal modulation format based on time-frequency analysis and denoising as claimed in claim 7, further comprising:

and the evaluation module is used for comparing the identification result with a real label to obtain the identification rate of the radar signal modulation format, and evaluating the identification performance of the improved Googlenet network according to the identification rate.

Images