Compressed residual network and radar radiation source signal identification method

Abstract

The invention belongs to the technical field of radar target identification and radiation source identification, and relates to a compressed residual network and a radar radiation source signal identification method. The method comprises the following steps: preprocessing a radar radiation source signal to obtain a preprocessed image; the preprocessing comprises CWD time-frequency transformation, graying and time-frequency image size adjustment; the preprocessed image is sent to a conv1, then maxpooling processing is carried out, and then the image passes through a conv2; and after the images sequentially pass through conv3 to conv6 and GAP, inputting the images processed by the GAP into softmax for classification, and outputting a classification result. Each of the conv2 to the conv6 uses a residual unit; and when the sizes and dimensions of the input and output feature maps are inconsistent, matching the input feature maps by adopting a 1 * 1 convolution kernel. The method has a good identification effect on the large-similarity radar radiation source.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target recognition and radiation source recognition, and relates to a compressed residual network (CRN) for radar radiation source signal recognition and a radar radiation source signal recognition method based on the compressed residual network. Background Art

[0002] At present, there are many types of radar systems and radar waveforms, and they have the characteristics of complex signal forms, surging signal density, and a large range of signal-to-noise ratio variations, which makes radar radiation source signal recognition challenging. The traditional manual intra-pulse analysis method has the technical defects of low recognition rate, poor stability, weak generalization ability, and almost inability to extract signal features under low signal-to-noise ratio conditions.

[0003] In order to accurately and effectively identify radar emitter signals and improve the intelligence level of electronic countermeasure equipment, the introduction of deep learning technology into radar emitter signal recognition has become a development trend in recent years. The convolutional neural network model automatically extracts features from data to complete model training by constructing a loss function, using gradient descent and error back propagation, and has become the main technical method for image classification, speech recognition, and target detection.

[0004] Time-frequency analysis is used to transform time-domain signals into time-frequency images, and the time-domain signal recognition problem is converted into image recognition, and then convolutional neural networks are used to complete the recognition of different radar emitter signals. In order to further improve the recognition rate, open source models such as AlexNet, GoogLeNet, and ResNet with strong performance and good generalization ability are used to complete the recognition of radar emitter signals with high recognition rate and good robustness.

[0005] However, existing recognition methods have the following problems:

[0006] 1) All of them use simple network models or open source network models. Simple network models have limited performance. Although complex open source network models can perform radar signal recognition, their design methods and application scope tend to be conventional image recognition. Therefore, the above scholars did not consider the characteristics of radar signal time-frequency graphs and tried to design a network model suitable for radar signal recognition that can balance recognition performance and complexity;

[0007] 2) The signal types are mostly 5 to 8, and the signal characteristics are obvious and relatively easy to identify. There are few studies on more than 10 types of signals, especially the analysis of easily confused signals.

[0008] Based on this, this application is committed to combining the characteristics of radar time-frequency images under low signal-to-noise ratio conditions, the recognition mechanism of convolutional neural networks, and exploring the recognition of radar emitter signals under low signal-to-noise ratio conditions, especially the recognition of multi-phase coded signals (FRANK, P1, P2, P3, P4) with similar time-frequency characteristics. Drawing on the design idea of ​​dilated residual network (DRN), a compressed residual network (CRN) and CRN model suitable for radar signal recognition are proposed and designed. The model has a suitable size and balances recognition performance and complexity. Summary of the invention

[0009] The purpose of the present invention is to address the defects of difficulty in radar emitter identification, high complexity of identification methods and low identification performance of multi-phase coded emitter signals (FRANK, P1, P2, P3, P4) with similar time-frequency characteristics under low signal-to-noise ratio conditions. Combined with the characteristics of radar time-frequency images, a radar emitter signal identification method based on a compressed residual network is proposed to identify radar emitter signals with high similarity.

[0010] In order to achieve the above object, the present invention adopts the following scheme.

[0011] The compressed residual network for radar emitter signal recognition includes conv1, maxpooling, conv2, conv3, conv4, conv5, conv6, GAP and softmax layers;

[0012] The conv1 is connected to the maxpooling with stride 2, the maxpooling is connected to conv2, the conv2 is connected to conv3, the conv3 is connected to conv4, the conv4 is connected to conv5, the conv5 is connected to conv6, the conv6 is connected to GAP, and the GAP is connected to the softmax layer;

[0013] The parameters of conv1 are: 7*7,conv,32;

[0014] The parameters of the conv2 are two 3*3,conv,32;

[0015] The parameters of conv3 are 3*3,conv,64;

[0016] The parameters of the conv4 are two 3*3,conv,128;

[0017] The parameters of the conv5 are two 3*3,conv,256;

[0018] The parameters of the conv6 are two 3*3,conv,256.

[0019] The conv in conv1, conv2, conv3, conv4, conv5 and conv6 is a normal convolution; one residual unit is used in each of conv2, conv3, conv4, conv5 and conv6.

[0020] When the input and output feature map sizes and dimensions are inconsistent, conv3, conv4, conv5, and conv6 use 1*1 convolution kernels to match the input feature maps.

[0021] The conv6 includes two models according to the degree of compression:

[0022] In the first model, the conv6 input feature map is not padded, and a 3*3 convolution operation is performed directly with a convolution step size of 1. Finally, the output feature map size is 5*5.

[0023] In the second model, a 3*3 convolution operation with a stride of 1 is directly performed on the conv6 input feature map, and then the stride is set to 2, resulting in an output feature map size of 3*3.

[0024] The radar emitter signal recognition method based on the compressed residual network comprises the following steps:

[0025] S1, preprocessing the radar radiation source signal to obtain a preprocessed image;

[0026] S2, send the preprocessed image to conv1, and output the processed image of conv1;

[0027] The parameters of conv1 are: 7*7,conv,32;

[0028] S3, perform maxpooling processing on the image processed by conv1 with a stride of 2 to obtain the processed image;

[0029] S4, input the processed image into conv2, and output the processed image of conv2;

[0030] The parameters of the conv2 are two 3*3,conv,32;

[0031] S5, input the image processed by conv2 into conv3, and output the image processed by conv3;

[0032] The parameters of conv3 are 3*3,conv,64;

[0033] S6, input the image processed by conv3 into conv4, and output the image processed by conv4;

[0034] The parameters of the conv4 are two 3*3,conv,128;

[0035] S7, input the image processed by conv4 into conv5, and output the image processed by conv5;

[0036] The parameters of the conv5 are two 3*3,conv,256;

[0037] S8, input the image processed by conv5 into conv6, and output the image processed by conv6;

[0038] The parameters of the conv6 are two 3*3,conv,256;

[0039] S9, input the image processed by conv6 into GAP, and output the image processed by GAP;

[0040] The GAP processed image is input into softmax,14 for classification, and the classification result is output. The preprocessing in S1 includes CWD time-frequency transformation, grayscale conversion and adjustment of the time-frequency image size; the CWD, namely Choi-Williams Distribution, represents Choi-Williams distribution.

[0041] S1, including the following sub-steps:

[0042] S11, performing CWD time-frequency transformation on various radar radiation source signals to obtain a transformed time-frequency diagram;

[0043] S12, graying the transformed time-frequency image to obtain a grayscale image;

[0044] The grayscale processing of S12 retains the complete information of the time-frequency image under various signal-to-noise ratio conditions;

[0045] S13, adjusting the grayscale image obtained in S12 to an image of the same size, to obtain a preprocessed image.

[0046] The conv in conv1, conv2, conv3, conv4, conv5 and conv6 is a normal convolution; one residual unit is used in each of conv2, conv3, conv4, conv5 and conv6.

[0047] When the input and output feature map sizes and dimensions are inconsistent, conv3, conv4, conv5, and conv6 use 1*1 convolution kernels to match the input feature maps.

[0048] The conv6 includes two models according to the degree of compression:

[0049] In the first model, the conv6 input feature map is not padded, and a 3*3 convolution operation is performed directly with a convolution step size of 1. Finally, the output feature map size is 5*5.

[0050] In the second model, a 3*3 convolution operation with a stride of 1 is directly performed on the conv6 input feature map, and then the stride is set to 2, resulting in an output feature map size of 3*3.

[0051] Beneficial Effects

[0052] Compared with the prior art, the present invention provides a compressed residual network and a radar emitter signal recognition method, which has the following beneficial effects:

[0053] 1. The method can effectively suppress the noise signal in the conv6 output feature map and highlight the feature signal that is crucial for classification and recognition;

[0054] 2. The method solves the problems of difficulty in identifying radar emitters of multi-phase coded emitter signals with similar time-frequency characteristics under low signal-to-noise ratio conditions, high complexity of the identification method and low identification performance;

[0055] 3. The method is effective in identifying radar radiation source signals with greater similarity. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a process of a radar radiation source signal recognition method based on a compressed residual network of the present invention;

[0057] Figure 2 It is a radar radiation source signal recognition method based on a compressed residual network of the present invention and a recognition rate variation curve of 14 kinds of signals based on design models based on different methods;

[0058] Figure 3 is the confusion matrix of different methods for identifying multi-phase encoding;

[0059] Figure 4 It is the recognition of 14 radar radiation source signals by three models;

[0060] Figure 5 is the recognition rate of different methods for multi-phase coded signals. DETAILED DESCRIPTION

[0061] In the following, a radar emitter signal recognition method based on a compressed residual network of the present invention is further illustrated and described in detail in conjunction with the accompanying drawings and embodiments.

[0062] Example 1

[0063] The present invention discloses a radar emitter signal recognition method based on a compressed residual network. The data set is constructed by setting parameters in two aspects to increase the signal recognition difficulty, thereby helping to highlight the superiority of the method of the present invention. First, the same parameters are set for signals with similar characteristics, such as polyphase coded signals, BPSK and QPSK signals. 14 radar emitter signals are simulated, with a pulse width of 2μs and a sampling frequency of f. s The frequency band is 400MHz, and the parameter settings of different signals are shown in Table 1, which are randomly selected within the setting range. Second, a lower signal-to-noise ratio is set to test the model's ability to recognize signals of poor quality. The signal-to-noise ratio range is -16 to 0dB, with a step size of 2dB. 600 pulse modulation signals are generated for each type of signal under each signal-to-noise ratio, and then CWD time-frequency analysis and data preprocessing are used to generate a data set. Among them, the sample ratio of the training set, validation set, and test set is 4:1:1. The training set and validation set are used to train the model. After the training is completed, the trained model is used to perform recognition tests on various signals in the test set.

[0064] Table 1 Signal parameter settings

[0065]

[0066]

[0067] In order to compare the CRN model proposed in this application with the DRN model proposed in the prior art in terms of recognition rate, timeliness, etc., two models were designed. The final convolution layer output feature map sizes were: 5*5 and 3*3, respectively. These are the two methods described in this application: Method 1 and 2. A DRN model of equal size is designed for the design method of radar emitter signal recognition based on the dilated residual network provided in the prior art. The specific method is to add a conv6 convolution structure on the basis of the model proposed in the prior art. The conv6 convolution part of this structure does not perform downsampling operations. Therefore, the feature map size remains unchanged, and the convolution kernel expansion factor d is 4, which is the same as conv5. The three models have the same size, the same convolution layer, activation layer, pooling layer, and batch normalization layer. The difference is that the DRN model designed in the prior art uses dilated convolution operations in the conv4~conv6 part, while the CRN model of this application not only does not perform dilation operations, but also compresses the feature map size that finally enters the GAP operation. The two are the same in convolution layer, activation layer, pooling layer, and batch normalization layer.

[0068] The GAP, or global average pooling, can effectively alleviate the overfitting problem in model training, thereby improving the classification performance of the model: In the formula, f is the output value after GAP operation, M and N are the feature map sizes, and x ij is the pixel value of the feature map.

[0069] For conventional images, the DRN model has a higher resolution, and the final output feature map contains more feature information, which is more helpful for classification after the GAP operation. For radar time-frequency images, when the signal-to-noise ratio is high and the image quality is good, although the feature information in the final output feature map is relatively small, the features are relatively obvious and the feature energy is high, which is also helpful for classification. When the signal-to-noise ratio is low, the feature information extracted by the DRN model is no longer obvious, and more subtle information is "submerged" by the noise. Increasing the feature resolution leads to limited extraction of useful features and a large increase in noise interference information. Therefore, the noise information with strong energy in the final output feature map increases significantly. After the GAP operation, the signal feature information is more likely to be "covered" by random noise information, which makes model classification difficult.

[0070] In the specific implementation, a data set is generated by using CWD time-frequency analysis and data preprocessing. The data set includes a training set, a validation set, and a test set. The training set and the validation set are used to train the CRN model. After the training is completed, the trained model is used to perform recognition tests on various types of signals in the test set.

[0071] The compressed residual network model (CRN model) includes conv1, maxpooling, conv2, conv3, conv4, conv5, conv6, GAP and softmax layers; conv1 is connected to maxpooling, maxpooling is connected to conv2, conv2 is connected to conv3, conv3 is connected to conv4, conv4 is connected to conv5, conv5 is connected to conv6, conv6 is connected to GAP, and GAP is connected to the softmax layer;

[0072] The parameters of conv1 are: 7*7,conv,32,

[0073] The parameters of the conv2 are two 3*3,conv,32;

[0074] The parameters of the conv3 are two 3*3,conv,64;

[0075] The parameters of the conv4 are two 3*3,conv,128;

[0076] The parameters of the conv5 are two 3*3,conv,256;

[0077] The parameters of the conv6 are two 3*3,conv,256;

[0078] The conv in conv1, conv2, conv3, conv4, conv5 and conv6 is a normal convolution; one residual unit is used in each of conv2, conv3, conv4, conv5 and conv6;

[0079] When the input and output feature map sizes and dimensions are inconsistent, conv3, conv4, conv5, and conv6 use 1*1 convolution kernels to match the input feature maps;

[0080] After each convolution operation, a batch normalization operation is performed, that is, a BN layer is added and Relu is used as the activation function after the BN layer;

[0081] The conv6 includes two models according to the degree of compression:

[0082] In the first model, the conv6 input feature map is not padded, and a 3*3 convolution operation is performed directly with a convolution step size of 1. Finally, the output feature map size is 5*5.

[0083] In the second model, a 3*3 convolution operation with a stride of 1 is directly performed on the conv6 input feature map, and then the stride is set to 2, resulting in an output feature map size of 3*3;

[0084] The convolutional layers, pooling layers, and batch normalization layers of the CRN are the same as those of the CNN and ResNet models;

[0085] The radar radiation source signal identification method has the following process: Figure 1 As shown, the following steps are included:

[0086] S1, preprocessing the radar radiation source signal to obtain a preprocessed image;

[0087] The preprocessing includes CWD time-frequency transformation, graying and adjusting the size of the time-frequency image; the CWD, namely Choi-Williams Distribution, represents Choi-Williams distribution;

[0088] S1, including the following sub-steps:

[0089] S11, performing CWD time-frequency transformation on various radar radiation source signals to obtain a transformed time-frequency diagram;

[0090] S12, graying the transformed time-frequency image to obtain a grayscale image;

[0091] The grayscale processing of S12 retains the complete information of the time-frequency image under various signal-to-noise ratio conditions;

[0092] S13, adjusting the grayscale image obtained in S12 to an image of the same size, to obtain a preprocessed image;

[0093] When S1 is implemented, CWD is transformed into:

[0094]

[0095] Among them, t and w represent time and angular frequency, σ is the attenuation coefficient. In order to take into account both high time-frequency resolution and suppression of cross terms, σ is set to 1.

[0096] S2, send the preprocessed image to conv1, and output the processed image of conv1;

[0097] The parameters of conv1 are: 7*7, conv, 32, stride2; where conv is a normal convolution;

[0098] S3, perform maxpooling processing on the image processed by conv1 to obtain the processed image;

[0099] S4, input the processed image into conv2, and output the processed image of conv2;

[0100] The parameters of the conv2 are two 3*3,conv,32;

[0101] S5, input the image processed by conv2 into conv3, and output the image processed by conv3;

[0102] The parameters of conv3 are 3*3,conv,32,stride2 and 3*3,conv,64;

[0103] S6, input the image processed by conv3 into conv4, and output the image processed by conv4;

[0104] The parameters of the conv4 are two 3*3,conv,128;

[0105] S7, input the image processed by conv4 into conv5, and output the image processed by conv5;

[0106] The parameters of the conv5 are two 3*3,conv,256;

[0107] S8, input the image processed by conv5 into conv6, and output the image processed by conv6;

[0108] The parameters of the conv6 are two 3*3,conv,256;

[0109] S9, input the image processed by conv6 into GAP, and output the image processed by GAP;

[0110] The GAP-processed image is input into softmax,14 for classification, and the classification result is output.

[0111] In order to compare the CRN model and recognition method proposed in the present invention with the existing DRN model, this embodiment is compared in terms of recognition rate and timeliness. Two models are designed based on the method, and the final convolution layer outputs feature map sizes of 5*5 and 3*3 respectively. This is the method 1 and 2. A DRN model of the same size is designed based on the method provided by the prior art. The specific method is to add a conv6_x convolution structure on the basis of the proposed model. The conv6_1 convolution part of the structure does not perform downsampling operation. Therefore, the feature map size remains unchanged, and the convolution kernel expansion factor d is 4, which is the same as conv5_x. The three models have the same size, the same convolution layer, activation layer, pooling layer, and batch normalization layer. The difference is that the DRN model provided by the prior art uses an expansion convolution operation in the conv4_x~conv6_x part, while the CRN model based on the present invention not only does not perform an expansion operation but also compresses the feature map size that finally enters the GAP operation.

[0112] Comparison of computational load: Table 1 shows the comparison of multiplication and addition computational load of CRN and DRN models. The present invention proposes two CRN models:

[0113] The conv6 includes two models according to the degree of compression:

[0114] In the first model, the conv6 input feature map is not padded, and a 3*3 convolution operation is performed directly with a convolution step size of 1. Finally, the output feature map size is 5*5.

[0115] In the second model, a 3*3 convolution operation with a stride of 1 is directly performed on the conv6 input feature map, and then the stride is set to 2, so that the output feature map size is 3*3; the multiplication and addition calculations required by the two models are the same for conv1~conv5, and the difference is only reflected in conv6. Overall, the calculation amount of the two models is similar, and the second model is selected for comparison with the model proposed by the prior art, that is, the model in which the feature map size is compressed to 3*3. The CRN and DRN models have the same conv1~conv3 structure and the same multiplication and addition calculation amount. The number of input and output channels of conv4~conv6 is relatively high. Since the DRN model expands the feature map size, the calculation amount of the underlying convolution part will increase exponentially. The CRN model not only does not expand the feature map size, but also performs further size compression in the last layer, which greatly reduces the multiplication and addition calculation amount of this part.

[0116] Timeliness comparison: Convolutional neural network models have achieved excellent results in fields such as image recognition, but their complex network scale and huge computational overhead limit their promotion and application in engineering. When designing a network model, we must consider both timeliness and model scale while considering recognition performance. The network scale is related to the number of model parameters, which is mainly determined by the number of parameters of the convolutional layer and the fully connected layer. Therefore, the CRN model proposed in this application and the DRN model of the prior art have the same number of parameters, so the model scale is similar. The measure of timeliness is the time required to complete an identification, which is mainly related to the amount of multiplication and addition calculations required for an identification. The amount of multiplication and addition calculations S required for the convolution operation can be expressed as:

[0117] S=D in *D in *M*N*D out *D out ;

[0118] Among them, D in Input size for feature map, D out is the feature map output size, M and N are the number of input and output feature channels respectively. Table 2 shows the output size of the feature map using the formula S = D in *D in *M*N*D out *D out The multiplication and addition calculations required for convolution operations at each layer of different models are calculated. For the sake of convenience, the point-by-point convolution calculations for dimension increase, which have relatively small calculations, are ignored. The comparison of the calculations of the second model and the prior art is shown in Table 2.

[0119] Table 2 The amount of multiplication and addition required for each convolution operation of different models

[0120]

[0121] Table 2 After calculation, the number of multiplication and addition calculations required for one image recognition is about 1912M for the DRN model proposed in the prior art, and about 218M for the CRN model of the present invention, which is only 11.4% of the DRN model. Therefore, the timeliness of the CRN model is much higher than that of the DRN model. The above different models were used to recognize 20,000 radar time-frequency images. The tests were all carried out in an environment where the CPU was AMD Ryzen55600X and the GPU was NAVIDIA GeForce RTX 3060. The average time for each model to perform one recognition is as follows: 0.43ms for CRN and 1.77ms for DRN model. The recognition timeliness of the CRN model is about four times that of the DRN model.

[0122] Recognition rate comparison: Figure 2 The following are the recognition rate change curves of 14 signals based on the design models of different methods. In general, the performance of the three models is relatively stable. As the signal-to-noise ratio increases, the recognition rate of the signal is steadily increasing. In contrast, the recognition rates of the two CRN models designed by the present invention are relatively close, and they are higher than the DRN model provided by the prior art under various signal-to-noise ratio conditions. After calculation, the average recognition rates of the CRN model of the present application are 93.01% and 93.03%, respectively, while the DRN model designed based on the prior art is only 84.08%. The designed CRN model reduces the influence of noise energy in the feature map by compressing the feature map size, making it easier to extract the feature information of the signal, and the model has good performance and a high recognition rate for various types of signals. The DRN model in the prior art increases the feature map size, which helps to extract more feature information, but the radar time-frequency map feature information is relatively small, so the added feature information is limited, and due to the low signal-to-noise ratio, a large amount of high-energy noise information is added, "flooding" the feature information of the signal, so the model performance is reduced, and the recognition rate for various types of signals is low.

[0123] In order to further compare the recognition performance of different methods for easily confused signals, Figure 3The confusion matrix of different methods for identifying polyphase coding when the signal-to-noise ratio is -14dB. (3a) is the proposed model, and (3b) is the method described in the prior art. Among them, the conv6 corresponding to the first model is: the conv6 input feature map is not padded, and a 3*3 convolution operation is directly performed, the convolution step stride is 1, and the output feature map size is 5*5; the conv6 input feature map corresponding to the second model directly performs a 3*3 convolution operation with a convolution step stride of 1, and then the stride is set to 2, and the output feature map size is 3*3. Each method has different degrees of confusion on the polyphase coded signal, especially P1, P2 and P4 codes, which have similar features and are very easy to be confused. The main difference between P1, P2 and P4 codes is the relatively weak "staircase" shape of the energy generated by the step frequency. The DRN provided by the prior art will introduce a large amount of noise energy under low signal-to-noise ratio conditions due to the expansion of the feature size, so that the feature map information used for classification is "flooded" by random noise. Therefore, the method described in the prior art can hardly identify P1 and P2 signals and mistakenly identifies them as P4 signals. The CRN model constructed by the method described in the present invention reduces the energy of random noise by compressing the feature map size finally used for classification, so that the model can extract effective feature map information and can identify P1 and P2 signals to a certain extent.

[0124] Depend on Figure 4 It can be seen that the three models are relatively stable in recognizing 14 radar emitter signals, and the recognition rate is steadily increasing with the improvement of the signal-to-noise ratio. Under various signal-to-noise ratio conditions, CRN performs best, followed by the ResNet model, and the lower the signal-to-noise ratio, the more obvious the superiority of the CRN model. This verifies the CRN model's advantage in recognizing radar emitter signals under low signal-to-noise ratio conditions: size compression reduces the energy information of random noise, making it easier to extract the feature information of the signal, making the effective information of the feature map used for classification more obvious, thereby achieving higher recognition performance.

[0125] In order to further compare the recognition performance of each method, the recognition rate of easily confused signals is analyzed. Figure 5 The recognition rates of different methods for polyphase coded signals are (5a) FRANK, (5b) P1, (5c) P2, (5d) P3, and (5e) P4. Since the characteristics of polyphase coded signals are similar and the simulation parameters are set the same, it can be seen from the time-frequency distribution diagram that P1 and P2 are both composed of a stepped oblique line, and the "staircase" shape is similar and almost indistinguishable to the naked eye; FRANK and P3 are both composed of two parallel lines, the difference is that FRANK has a certain "staircase" shape; P4 code is composed of a parallel line, similar to P1 and P2. Therefore, the above signals are very easy to confuse under low signal-to-noise ratio conditions. Whether the easily confused signals can be correctly identified determines the overall recognition performance of the model.

[0126] Example 2

[0127] The second embodiment of the present invention relates to a radar emitter signal identification device based on a compressed residual network, comprising at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned radar emitter signal identification method based on a compressed residual network (Compressed Residual Network, CRN).

[0128] Among them, the memory and the processor are connected in a bus manner, and the bus may include any number of interconnected buses and bridges, and the bus connects various circuits of one or more processors and memories together. The bus can also connect various other circuits such as peripheral devices, voltage regulators and power management circuits through interfaces, which are well known in the art. The interface provides an interface between the bus and the transceiver, such as a communication interface and a user interface. The transceiver can be one element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices on a transmission medium. The data processed by the processor is transmitted on a wireless medium through an antenna, and further, the antenna also receives data and transmits the data to the processor. The processor is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management and other control functions. The memory can be used to store data used by the processor when performing operations.

[0129] Example 3

[0130] The third embodiment of the present invention relates to a computer-readable storage medium storing a computer program, and the computer program implements the above method embodiment when executed by a processor.

[0131] Those skilled in the art can understand from the above description that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program is stored in a storage medium, including several instructions for making a device (which can be a single-chip microcomputer, chip, etc.) or a processor (processor) execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes but is not limited to various media that can store program codes, such as USB flash drives, mobile hard disks, magnetic storage devices, and optical storage devices.

[0132] It can be seen from the above description that the embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. The above is only a preferred embodiment of the present invention, and the present invention should not be limited to the contents disclosed in the embodiment and the accompanying drawings. Any equivalent or modification that is completed without departing from the spirit disclosed in the present invention falls within the scope of protection of the present invention.

Claims

1. A compressed residual network for radar emitter signal recognition, characterized in that: Including conv1, maxpooling with stride 2, conv2, conv3, conv4, conv5, conv6, GAP and softmax layers; The conv1 is connected to the maxpooling with stride 2, the maxpooling is connected to conv2, the conv2 is connected to conv3, the conv3 is connected to conv4, the conv4 is connected to conv5, the conv5 is connected to conv6, the conv6 is connected to GAP, and the GAP is connected to the softmax layer; The conv in conv1, conv2, conv3, conv4, conv5 and conv6 are all ordinary convolutions; one residual unit is used in conv2, conv3, conv4, conv5 and conv6.

2. The compressed residual network according to claim 1, characterized in that When the input and output feature map sizes and dimensions are inconsistent, conv3, conv4, conv5, and conv6 use 1*1 convolution kernels to match the input feature maps.

3. The compressed residual network according to claim 1, characterized in that The conv6 includes two models according to the degree of compression: In the first model, the conv6 input feature map is not padded, and a 3*3 convolution operation is performed directly with a convolution step size of 1. Finally, the output feature map size is 5*5. In the second model, a 3*3 convolution operation with a stride of 1 is directly performed on the conv6 input feature map, and then the stride is set to 2, resulting in an output feature map size of 3*3.

4. The compressed residual network according to claim 1, characterized in that The parameters of the conv1 are: 7*7, conv,32; the parameters of the conv2 are two 3*3, conv,32; the parameters of the conv3 are two 3*3, stride is 2, conv,64; the parameters of the conv4 are two 3*3, stride is 2, conv,128; the parameters of the conv5 are two 3*3, stride is 2, conv,256; the parameters of the conv6 are two 3*3, conv,256.

5. A radar emitter signal recognition method based on a compressed residual network, characterized in that: The steps include: S1, preprocessing the radar radiation source signal to obtain a preprocessed image; The preprocessing includes CWD time-frequency transformation, graying and adjusting the size of the time-frequency image; the CWD, namely Choi-Williams Distribution, represents Choi-Williams distribution; S2, send the preprocessed image to conv1, and output the processed image of conv1; S3, perform maxpooling processing on the image processed by conv1 to obtain the processed image; S4, input the processed image into conv2, and output the processed image of conv2; S5, input the image processed by conv2 into conv3, and output the image processed by conv3; S6, input the image processed by conv3 into conv4, and output the image processed by conv4; S7, input the image processed by conv4 into conv5, and output the image processed by conv5; S8, input the image processed by conv5 into conv6, and output the image processed by conv6; S9, input the image processed by conv6 into GAP, and output the image processed by GAP; Input the GAP-processed image into softmax for classification and output the classification result; The conv in the conv1, conv2, conv3, conv4, conv5 and conv6 is a normal convolution and one residual unit is used in each of the conv2, conv3, conv4, conv5 and conv6.

6. The radar radiation source signal identification method according to claim 5, characterized in that: S1, including the following sub-steps: S11, performing CWD time-frequency transformation on various radar radiation source signals to obtain a transformed time-frequency diagram; S12, graying the transformed time-frequency image to obtain a grayscale image; S13, adjusting the grayscale image obtained in S12 to an image of the same size, to obtain a preprocessed image.

7. The radar radiation source signal identification method according to claim 6, characterized in that: The stride of the maxpooling process in S3 is 2; the grayscale process in S12 retains the complete information of the time-frequency image under various signal-to-noise ratio conditions.

8. The radar radiation source signal identification method according to claim 5, characterized in that: When the input and output feature map sizes and dimensions are inconsistent, conv3, conv4, conv5, and conv6 use 1*1 convolution kernels to match the input feature maps.

9. The radar radiation source signal identification method according to claim 5, characterized in that: The conv6 includes two models according to the degree of compression: In the first model, the conv6 input feature map is not padded, and a 3*3 convolution operation is performed directly with a convolution step size of 1. Finally, the output feature map size is 5*5. In the second model, a 3*3 convolution operation with a stride of 1 is directly performed on the conv6 input feature map, and then the stride is set to 2, so that the output feature map size is 3*3. The parameters of the conv1 are: 7*7,conv,32.

10. The radar radiation source signal identification method according to claim 5, characterized in that: The parameters of the conv2 are two 3*3, conv,32; the parameters of the conv3 are 3*3, conv,64; the parameters of the conv4 are two 3*3, conv,128; the parameters of the conv5 are two 3*3, conv,256; the parameters of the conv6 are two 3*3, conv,256.

Images