Radar Signal Modulation Recognition Method Based on Convolutional Denoising and Resnet50

Abstract

The present invention discloses a radar signal modulation recognition method based on convolutional denoising and Resnet50. The implementation scheme of the present invention is as follows: perform time-frequency analysis on radar signals; generate a training set; construct a convolutional denoising autoencoder network; train the convolutional denoising autoencoder network; construct a deep residual neural network 19-Resnet50; use a transfer learning strategy to train the deep residual neural network 19-Resnet50; identify the in-pulse modulation type of radar signals. The present invention uses a convolutional denoising autoencoder network to realize the denoising of the time-frequency image of radar signals under low signal-to-noise ratio, and improves the quality of the time-frequency feature image of radar signals; the present invention uses a deep residual neural network 19-Resnet50 to extract deeper image classification features, and can identify the in-pulse modulation types of various radar signals in a low signal-to-noise ratio environment, having the advantages of strong anti-noise ability, a large number of recognized radar signal types, and high recognition accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar communication, and further relates to a method for identifying the intra-pulse modulation of radar signals based on convolutional denoising processing and deep residual neural network 19-Resnet50 in the field of electronic countermeasures. The present invention can be used to effectively identify the intra-pulse modulation type of radar signals in an environment with low signal-to-noise ratio and multiple radar signal modulation categories. Background Art

[0002] The identification of the intra-pulse modulation type of radar signals is an important part of electronic reconnaissance. As a key technology in radar countermeasures, it has important applications in radar reconnaissance systems. Using the intra-pulse modulation type of radar signals as a sorting parameter is an effective way to improve the sorting and recognition rate of radar signals. The intra-pulse modulation type of radar signals is also an important parameter for radar model identification, threat level assessment, and interference decision-making. With the development of radar technology, lower signal-to-noise ratio and more radar signal modulation categories have made the identification of the intra-pulse modulation type of radar signals more difficult. Therefore, the technology for identifying the intra-pulse modulation type of radar signals has received more attention in the field of electronic reconnaissance.

[0003] Shi Limeng et al. disclosed a method for identifying radar signal modulation using a recurrent neural network and a convolutional neural network to extract the time-domain and time-frequency domain features of radar signals respectively in their published paper "Identification of Radar Signal Modulation Modes Based on Deep Networks" (Journal of Ordnance Equipment Engineering, 2021, 42(6): 190-193+218). The implementation scheme of this method is as follows: First step, perform time-frequency analysis on radar signals: Use the quadratic time-frequency analysis method Choi-Williams distribution to perform time-frequency analysis to obtain the time-frequency image of radar signals. Second step, construct a long short-term memory network and a convolutional neural network model: Use the long short-term memory network to extract the time-domain features of radar signals, and use the convolutional neural network to extract the time-frequency domain features of radar signals. Third step, feature synthesis and recognition: Synthesize the extracted time-domain and time-frequency domain features to form joint features, and achieve classification and recognition through a fully connected neural network. This method can achieve the identification of common radar signal modulation types. However, the deficiencies of this method still lie in that the signal feature contamination by noise is serious under low signal-to-noise ratio in the identified radar signal modulation categories, resulting in limited radar signal modulation recognition categories.

[0004] Nanjing University of Aeronautics and Astronautics proposed a method for identifying the intra-pulse modulation of radar signals that combines time-frequency analysis and machine learning in its patent document "Method for Identifying Radar Signal Modulation Based on Time-Frequency Analysis and Machine Learning" (Application No.: CN 202110092719.2; Publication No.: CN 112560803 A). The implementation scheme of this method is as follows: In the first step, polyphase filtering is used to detect pulses in the pulse stream, and then the Choi-Williams analysis is performed on the detected pulse signals to obtain the time-frequency image of the radar signal. At the same time, the time-frequency image is binarized and denoised by morphology; In the second step, a residual neural network is constructed to perform an initial identification on the preprocessed time-frequency image to obtain the identification results other than polyphase codes; In the third step, the time-frequency image of the polyphase code is subjected to feature extraction again. After obtaining the feature parameters, they are fed into a random forest classifier for training to obtain the identification results of the polyphase code signal. This method can realize the identification of the modulation types of various radar signals including polyphase codes and multi-time codes, and has a good recognition rate at a low signal-to-noise ratio. However, the deficiencies of this method are still that the denoising process of the noisy time-frequency image at a low signal-to-noise ratio fails to well preserve the time-frequency features of the signal, and only using an 18-layer residual neural network results in insufficient time-frequency features being extracted, leading to a low recognition rate and fewer types of radar signals recognized at a low signal-to-noise ratio.

[0005] In summary, due to the poor extraction effect of the time-frequency features of radar signals at a low signal-to-noise ratio and the insufficient depth of the constructed feature extraction network, the recognition of the intra-pulse modulation types of radar signals is not ideal under the existing methods. Summary of the Invention

[0006] The purpose of the present invention is to propose a method for identifying the intra-pulse modulation of radar signals based on convolutional denoising processing and Resnet50 deep residual neural network in view of the above deficiencies of the existing technology, so as to solve the problems of poor denoising effect of noisy time-frequency images at a low signal-to-noise ratio, insufficient ability of the network to extract signal features, resulting in low recognition rate and few types of recognized radar signal modulation types at a low signal-to-noise ratio.

[0007] The specific idea for realizing the purpose of the present invention is: The present invention uses a convolutional denoising autoencoder network to denoise the time-frequency image of radar signals at a low signal-to-noise ratio. Compared with the traditional binarization and morphological image processing methods, the method of the present invention can make the noisy time-frequency image better fit the noise-free time-frequency image, thus solving the problem of poor feature extraction effect of the time-frequency image at a low signal-to-noise ratio. The present invention uses a deep residual neural network to solve the problem of network performance degradation caused by the increase in the number of layers of traditional convolutional neural networks, and uses a transfer learning strategy to train the deep residual neural network Resnet50, so as to solve the problems of low recognition rate and few types of recognized radar signal modulation types at a low signal-to-noise ratio.

[0008] To achieve the above object, the technical solution adopted by the invention includes the following steps:

[0009] Step 1, perform time-frequency analysis on the radar signal:

[0010] Step 1.1, perform Choi-Williams time-frequency analysis on 19 types of radar signals respectively to obtain the time-frequency image matrix corresponding to this type of radar signal;

[0011] Step 1.2, perform normalization processing on each time-frequency image matrix;

[0012] Step 2, generate a training set:

[0013] Step 2.1, select at least 240 time-frequency images of each type of radar signal after normalization processing. Form a sample set with 120 noisy time-frequency images among them, and form a sample label set with the corresponding 120 noise-free time-frequency images. Combine all samples in the sample set and their corresponding labels to form training set 1;

[0014] Step 2.2, select at least 200 noisy time-frequency images of each type of radar signal after normalization processing to form a sample set, and use the image category as the sample label set. Combine all samples in the sample set and their corresponding labels to form training set 2;

[0015] Step 3, construct a convolutional denoising autoencoder network:

[0016] Step 3.1, construct the encoder in the convolutional denoising autoencoder network. Its structure is connected in series in turn as: the first convolutional layer, ReLu activation layer, the second convolutional layer, ReLu activation layer, the first pooling layer, the third convolutional layer, ReLu activation layer, the fourth convolutional layer, ReLu activation layer, the fifth convolutional layer, ReLu activation layer, the second pooling layer; set the number of convolutional kernels of the first to fifth convolutional layers to 64, 64, 128, 128, 256 in turn, set the size of the convolutional kernels to 3×3, both the first and second pooling layers use maximum pooling, set the size of the pooling kernels to 1×2, and set the pooling stride to 1×2;

[0017] Step 3.2, construct the decoder in the convolutional denoising autoencoder network. Its structure is connected in series in turn as: the first deconvolutional layer, ReLu activation layer, the second deconvolutional layer, ReLu activation layer, the third deconvolutional layer, ReLu activation layer, the fourth deconvolutional layer, ReLu activation layer, the fifth deconvolutional layer, ReLu activation layer, the sixth deconvolutional layer, ReLu activation layer, the seventh deconvolutional layer, ReLu activation layer; set the number of convolutional kernels of the first to seventh convolutional layers to 128, 128, 64, 32, 32, 16, 3 in turn, and set the size of the deconvolutional kernels to 3×3;

[0018] Step 3.3, connect the constructed encoder, decoder, and Sigmoid activation layer in series in sequence to form a convolutional denoising autoencoder network. The constructed convolutional denoising autoencoder network has a total of 27 layers;

[0019] Step 4, train the convolutional denoising autoencoder network:

[0020] Input the training set 1 into the convolutional denoising autoencoder network, and use the backpropagation gradient descent method to iteratively update the parameters of each layer of the network until the loss function of the network converges, obtaining a trained convolutional denoising autoencoder network;

[0021] Step 5, construct the deep residual neural network 19-Resnet50:

[0022] Construct a deep residual neural network 19-Resnet50 composed of a residual neural network, a fully connected layer, and a Softmax activation layer connected in series; set the number of output neurons of the fully connected layer to 19;

[0023] Step 6, use the transfer learning strategy to train the deep residual neural network 19-Resnet50:

[0024] Step 6.1, input the training set 2 into the trained convolutional denoising autoencoder network, and output the denoised time-frequency image;

[0025] Step 6.2, input the denoised time-frequency image into the deep residual neural network 19-Resnet50, use the transfer learning strategy, and use the backpropagation gradient descent method to iteratively update the parameters of each layer of the network until the cross-entropy loss function of the network converges, obtaining a trained deep residual neural network 19-Resnet50;

[0026] Step 7, identify the in-pulse modulation type of the radar signal:

[0027] Step 7.1, adopt the same method as in Step 1, perform time-frequency analysis and normalization processing on all samples to be identified, and input the normalized samples into the trained convolutional denoising autoencoder network, and output the denoised time-frequency image;

[0028] Step 7.2, input the denoised time-frequency image into the deep residual neural network model 19-Resnet50 for classification and recognition, and output the in-pulse modulation type of the radar signal.

[0029] The present invention has the following advantages compared with the prior art:

[0030] First, the present invention uses a convolutional denoising autoencoder network to achieve denoising of radar signal time-frequency images under low signal-to-noise ratios, overcoming the deficiency of poor feature extraction effect of time-frequency images in the prior art under low signal-to-noise ratios, improving the quality of radar signal time-frequency feature images, and enabling the present invention to identify and classify the intra-pulse modulation types of radar signals in a low signal-to-noise ratio environment.

[0031] Second, the present invention uses a transfer learning strategy to train a deep residual neural network 19-Resnet50, thereby overcoming the defect of network degradation that occurs when the number of network layers of a convolutional neural network increases. Therefore, deeper network layers can be fully utilized to extract deeper image classification features, enabling the present invention to have a high recognition accuracy in identifying various intra-pulse modulation types of radar signals in a low signal-to-noise ratio environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flowchart of the present invention;

[0033] Figure 2 are the original time-frequency diagrams and denoised time-frequency diagrams of 19 types of radar signals at a signal-to-noise ratio of -5 dB;

[0034] Figure 3 is a comparison diagram of the simulation effects of the present invention; wherein, Figure 3 (a) is a graph of recognition accuracy at different signal-to-noise ratios, Figure 3 (b) and Figure 3 (c) are schematic diagrams of confusion matrices at signal-to-noise ratios of -5 dB and 0 dB, respectively. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] To more clearly illustrate the present invention, the following further detailed description is made in conjunction with embodiments and the drawings.

[0036] Referring to Figure 1 , a further detailed description is made of the specific implementation steps of the embodiments of the present invention.

[0037] Step 1, perform time-frequency analysis on the radar signal.

[0038] The embodiments of the present invention select 19 types of radar signals, namely linear frequency modulation signals (LFM), frequency diversity signals (FreqDiversity), frequency coding signals (FSK), non-linear frequency modulation signals (QFM, SFM, TFM), frequency hopping signals (Costas), polyphase code signals (Frank, P1, P2, P3, P4), multi-time code signals (T1, T2, T3, T4), and composite modulation signals (LFM + BPSK, FSK + BPSK).

[0039] Step 1.1, perform Choi-Williams time-frequency analysis on 19 types of radar signals respectively to obtain the time-frequency image matrix corresponding to this type of radar signal.

[0040] Step 1.2, according to the following formula, perform normalization processing on each time-frequency image matrix respectively. The normalization conversion function is:

[0041]

[0042] where, X n ′ represents the nth image matrix after normalization processing, and X n represents the nth image matrix, and max(X n ) represents finding the maximum value of the nth image matrix X n .

[0043] Step 2, generate the training set.

[0044] Step 2.1, select at least 240 time-frequency images of each type of radar signal after normalization processing. Form a sample set with 120 noisy time-frequency images among them, and form a sample label set with their corresponding 120 noise-free time-frequency images. Combine all samples in the sample set and their corresponding labels to form training set 1.

[0045] Step 2.2, select at least 200 noisy time-frequency images of each type of radar signal after normalization processing to form a sample set, and use the image category as the sample label set. Combine all samples in the sample set and their corresponding labels to form training set 2.

[0046] Step 3, construct a convolutional denoising autoencoder network.

[0047] Step 3.1, construct the encoder in the convolutional denoising autoencoder network. Its structure is connected in series in turn as: the first convolutional layer, ReLu activation layer, the second convolutional layer, ReLu activation layer, the first pooling layer, the third convolutional layer, ReLu activation layer, the fourth convolutional layer, ReLu activation layer, the fifth convolutional layer, ReLu activation layer, the second pooling layer; set the number of convolutional kernels of the first to fifth convolutional layers to 64, 64, 128, 128, 256 in turn, set the size of the convolutional kernels to 3×3, both the first and second pooling layers use maximum pooling, set the size of the pooling kernels to 1×2, and set the pooling stride to 1×2.

[0048] Step 3.2, construct the decoder in the convolutional denoising autoencoder network. Its structure is successively connected in series as follows: the first transposed convolutional layer, ReLu activation layer, the second transposed convolutional layer, ReLu activation layer, the third transposed convolutional layer, ReLu activation layer, the fourth transposed convolutional layer, ReLu activation layer, the fifth transposed convolutional layer, ReLu activation layer, the sixth transposed convolutional layer, ReLu activation layer, the seventh transposed convolutional layer, ReLu activation layer; set the number of convolutional kernels of the first to seventh convolutional layers to 128, 128, 64, 32, 32, 16, 3 in sequence, and set the size of the transposed convolutional kernels to 3×3.

[0049] Step 3.3, connect the constructed encoder, decoder and Sigmoid activation layer in series in sequence to form a convolutional denoising autoencoder network. The constructed convolutional denoising autoencoder network has a total of 27 layers. The convolutional denoising autoencoder first constructs an encoder through a convolutional network to extract the deep features of the noisy time-frequency image, and then constructs a decoder through the transposed convolutional layer to restore the original features of the image and recover the original size of the image.

[0050] Step 4, train the convolutional denoising autoencoder network.

[0051] Step 4.1, input the sample data in training set 1 into the convolutional denoising autoencoder network. Set the initial learning rate to 1e-3, adopt the Adam function with the ability of adaptive learning rate optimization as the optimizer, set the batch size to 16, and use the backpropagation gradient descent method to iteratively update the parameters of each layer of the network until the loss function of the network converges, so as to obtain a trained convolutional denoising autoencoder network model.

[0052] The loss function is as follows:

[0053]

[0054] Among them, n represents the batch size in batch normalization processing, y k represents the k-th image output by the convolutional denoising autoencoder network, x k represents the k-th noisy image input into the convolutional denoising autoencoder network, ||·|| F is the F-norm operation.

[0055] In the embodiment of the present invention, use the trained convolutional denoising autoencoder network to perform denoising processing on the noisy time-frequency images of 19 types of radar signals, and output the denoised time-frequency images. Figure 2 (a) to Figure 2(s) are the time-frequency images of the chirp signal, frequency diversity signal, frequency coding signal, quadratic frequency modulation signal, sinusoidal frequency modulation signal, tangent frequency modulation signal, Costas frequency hopping signal, Frank signal, P1 signal, P2 signal, P3 signal, P4 signal, T1 signal, T2 signal, T3 signal, T4 signal, LFM + BPSK composite modulation signal, and FSK + BPSK composite modulation signal at a signal-to-noise ratio of -5 dB, respectively. Figure 2 (a) to Figure 2 (s) For each group of images, the left side is the time-frequency image with noise, and the right side is the time-frequency image after denoising. By combining the time-frequency image with noise and the time-frequency image after denoising, it can be seen that most of the noise in the time-frequency image with noise is eliminated in each group of time-frequency images after denoising, and at the same time, the original time-frequency characteristics of the signal are restored, achieving a good denoising effect.

[0056] Step 5, construct the deep residual neural network 19-Resnet50.

[0057] Construct a deep residual neural network 19-Resnet50 composed of a residual neural network, a fully connected layer, and a Softmax activation layer in series. The structure of the residual neural network adopts the residual neural network in the existing Resnet50 network, and the number of output neurons in the fully connected layer is 19.

[0058] Step 6, use the transfer learning strategy to train the deep residual neural network 19-Resnet50.

[0059] Step 6.1, input the training set 2 into the trained convolutional denoising autoencoder network, and output the time-frequency image after denoising.

[0060] Step 6.2, input the time-frequency image after denoising into the deep residual neural network 19-Resnet50. Using the transfer learning strategy, the initial learning rate is set to 0.01 during training. The cross-entropy function is used as the loss function of the network, the SGD function is used as the optimizer, the batch size is 8, and the backpropagation gradient descent method is used to iteratively update the parameters of each layer of the network until the loss function of the network converges, obtaining the trained deep residual neural network 19-Resnet50.

[0061] The transfer learning strategy is as follows: First, train the replaced fully connected neural network in the early stage of training. The network parameters in the residual neural network of the 19-Resnet50 network remain unchanged, and the loaded model parameters are the parameters obtained after the 19-Resnet50 network converges in the ImageNet dataset training. After the model converges, then train the entire deep residual neural network 19-Resnet50 to update the parameters.

[0062] Step 7, identify the intra-pulse modulation type of the radar signal.

[0063] Step 7.1, using the same method as in Step 1, perform time-frequency analysis and normalization on all samples to be identified, input the normalized samples into the trained convolutional denoising autoencoder network, and output the denoised time-frequency image.

[0064] Step 7.2, input the denoised time-frequency image into the deep residual neural network model 19-Resnet50 for classification and identification, and output the intra-pulse modulation type of the radar signal.

[0065] The following further illustrates the effect of the present invention in combination with simulation experiments:

[0066] 1. Simulation experiment conditions:

[0067] The hardware platform for the simulation experiment of the present invention is: the processor is an Intel i7 11800H CPU with a main frequency of 2.3 GHz and a memory of 16 GB.

[0068] The software platform for the simulation experiment of the present invention is: Windows 10 operating system, MATLAB 2018b, and python3.7.

[0069] The modulation parameter settings of the 19 types of radar signals used in the simulation experiment of the present invention are shown in Table 1:

[0070] Table 1. List of modulation parameters of 19 types of radar signals

[0071]

[0072]

[0073] Set the sampling rate f s to 200 MHz, the signal length to a random value between 256 and 512, set the modulation parameters of the radar signal according to Table 1, the signal-to-noise ratio ranges from -5 dB to 10 dB with an interval of 5 dB, the noise is additive white Gaussian noise, 300 time-frequency images are generated for each type of radar signal at each signal-to-noise ratio, a total of 22,800 time-frequency images, the training set contains 15,200 time-frequency images, the test set contains 7,600 time-frequency images, and the test set is used to test the recognition effect of the radar signal modulation type of the present invention.

[0074] 2. Simulation content and its result analysis:

[0075] The simulation experiment of the present invention uses the method of the present invention under the above simulation conditions to test the recognition effect of 19 types of radar signal modulation types, and the obtained recognition accuracy test results are as Figure 3As shown in (a), the schematic diagrams of the confusion matrices formed by the experimental test results at signal-to-noise ratios of -5 dB and 0 dB are respectively Figure 3 (b) and Figure 3 (c).

[0076] The effects of the present invention will be further described below in conjunction with the Figure 3 simulation diagram.

[0077] Figure 3 (a) is a line graph showing the recognition accuracy rates of the intrapulse modulation types of 19 types of radar signals under signal-to-noise ratios of -5 dB, 0 dB, 5 dB, and 10 dB respectively by using the method of the present invention. Figure 3 (b) is the confusion matrix obtained by statistically analyzing the experimental test results under a signal-to-noise ratio of -5 dB by using the method of the present invention. Figure 3 (c) is the confusion matrix obtained by statistically analyzing the experimental test results under a signal-to-noise ratio of 0 dB by using the method of the present invention.

[0078] It can be seen from Figure 3 (a) that as the signal-to-noise ratio increases, the recognition accuracy rate of the intrapulse modulation types of 19 types of radar signals by using the method of the present invention also increases. The recognition accuracy rate reaches 91.2% under the condition of a signal-to-noise ratio of -5 dB; when the signal-to-noise ratio is 0 dB, the recognition accuracy rate increases to 98.4%, proving that the method of the present invention has an ideal recognition effect on the intrapulse modulation types of radar signals in a low signal-to-noise ratio environment.

[0079] It can be seen from Figure 3 (b) that when the signal-to-noise ratio is -5 dB, the recognition accuracy rates of the method of the present invention for most of the intrapulse modulation types of radar signals all reach more than 95%. For polyphase codes (Frank, P1, P3, P4) and linear frequency modulation signals (LFM), the T2 and T4 signals in multi-time codes are prone to confusion because the time-frequency characteristics of these types of radar signals are similar. Therefore, in a low signal-to-noise ratio environment, the image features are lost due to the influence of noise, resulting in a poor recognition effect on these types of signals.

[0080] It can be seen from Figure 3 (c) that as the signal-to-noise ratio increases, in a 0 dB signal-to-noise ratio environment, the recognition accuracy rates of the method of the present invention for the radar signal modulation types except for the P4 signal all reach more than 96%, proving that the method of the present invention also has a good performance for the signal types that are difficult to distinguish among the 19 types of radar signals at low signal-to-noise ratios.

[0081] The above simulation experiment results show that: by combining the convolutional denoising autoencoder, the deep residual neural network, and the transfer learning strategy, the method of the present invention can identify the in-pulse modulation types of radar signals in an environment with low signal-to-noise ratio and a variety of radar signal types, with good performance and a high recognition accuracy.

Claims

1. A radar signal modulation recognition method based on convolutional denoising and Resnet50, characterized in that Denoise the noisy time-frequency image of the radar signal under low signal-to-noise ratio by using the trained convolutional denoising autoencoder network, and extract the classification features of the time-frequency image after denoising by using the trained deep residual neural network 19-Resnet50; the specific steps of this method are as follows: Step 1, perform time-frequency analysis on the radar signal: Step 1.1, perform Choi-Williams time-frequency analysis on 19 types of radar signals respectively to obtain the time-frequency image matrix corresponding to this type of radar signal; Step 1.2, perform normalization processing on each time-frequency image matrix; Step 2, generate the training set: Step 2.1, select at least 240 time-frequency images of each type of radar signal after normalization processing, form a sample set with 120 noisy time-frequency images among them, form a sample label set with the corresponding 120 noise-free time-frequency images, and form training set 1 with all samples in the sample set and their corresponding labels; Step 2.2, select at least 200 noisy time-frequency images of each type of radar signal after normalization processing to form a sample set, use the image category as the sample label set, and form training set 2 with all samples in the sample set and their corresponding labels; Step 3, construct the convolutional denoising autoencoder network: Step 3.1, construct the encoder in the convolutional denoising autoencoder network, and its structure is connected in series in turn as: the first convolutional layer, ReLu activation layer, the second convolutional layer, ReLu activation layer, the first pooling layer, the third convolutional layer, ReLu activation layer, the fourth convolutional layer, ReLu activation layer, the fifth convolutional layer, ReLu activation layer, the second pooling layer; set the number of convolutional kernels of the first to fifth convolutional layers to 64, 64, 128, 128, and 256 in turn, set the size of the convolutional kernel to 3×3, both the first and second pooling layers use max pooling, set the size of the pooling kernel to 1×2, and set the pooling stride to 1×2; Step 3.2, construct the decoder in the convolutional denoising autoencoder network, and its structure is connected in series in turn as: the first deconvolutional layer, ReLu activation layer, the second deconvolutional layer, ReLu activation layer, the third deconvolutional layer, ReLu activation layer, the fourth deconvolutional layer, ReLu activation layer, the fifth deconvolutional layer, ReLu activation layer, the sixth deconvolutional layer, ReLu activation layer, the seventh deconvolutional layer, ReLu activation layer; set the number of convolutional kernels of the first to seventh convolutional layers to 128, 128, 64, 32, 32, 16, and 3 in turn, and set the size of the deconvolutional kernel to 3×3; Step 3.3, connect the constructed encoder, decoder and Sigmoid activation layer in series in turn to form a convolutional denoising autoencoder network, and the constructed convolutional denoising autoencoder network has a total of 27 layers; Step 4, train the convolutional denoising autoencoder network: Input training set 1 into the convolutional denoising autoencoder network, use the backpropagation gradient descent method to iteratively update the parameters of each layer of the network until the loss function of the network converges, and obtain the trained convolutional denoising autoencoder network; Step 5, construct the deep residual neural network 19-Resnet50: Construct a deep residual neural network 19-Resnet50 composed of a residual neural network, a fully connected layer, and a Softmax activation layer connected in series; set the number of output neurons in the fully connected layer to 19; Step 6, train the deep residual neural network 19-Resnet50 using the transfer learning strategy: Step 6.1, input the training set 2 into the trained convolutional denoising autoencoder network, and output the denoised time-frequency image; Step 6.2, input the denoised time-frequency image into the deep residual neural network 19-Resnet50, use the transfer learning strategy, and use the backpropagation gradient descent method to iteratively update the parameters of each layer of the network until the cross-entropy loss function of the network converges, obtaining the trained deep residual neural network 19-Resnet50; Step 7, identify the in-pulse modulation type of the radar signal: Step 7.1, adopt the same method as in Step 1, perform time-frequency analysis and normalization processing on all samples to be identified, and input the normalized samples into the trained convolutional denoising autoencoder network to output the denoised time-frequency image; Step 7.2, input the denoised time-frequency image into the deep residual neural network model 19-Resnet50 for classification and recognition, and output the in-pulse modulation type of the radar signal.

2. The radar signal modulation recognition method based on convolutional denoising and Resnet50 according to claim 1, characterized in that The 19 types of radar signals described in Step 1.1 refer to linear frequency modulation signals, frequency diversity signals, frequency coding signals, quadratic frequency modulation signals, sinusoidal frequency modulation signals, tangent frequency modulation signals, Costas frequency hopping signals, Frank signals, P1 signals, P2 signals, P3 signals, P4 signals, T1 signals, T2 signals, T3 signals, T4 signals, LFM+BPSK composite modulation signals, and FSK+BPSK composite modulation signals.

3. The radar signal modulation recognition method based on convolutional denoising and Resnet50 according to claim 1, characterized in that, The normalization processing described in Step 1.2 is implemented by the following formula: Among them, X n ′ represents the nth image matrix after normalization processing, and X n represents the nth image matrix, and max() represents finding the maximum value of the matrix.

4. The radar signal modulation recognition method based on convolutional denoising and Resnet50 according to claim 1, wherein The loss function of the convolutional denoising autoencoder network described in Step 4 is as follows: Among them, n represents the batch size in batch normalization, and y k represents the k-th image output by the convolutional denoising autoencoder network, and x k represents the k-th noisy image input to the convolutional denoising autoencoder network, and ||·|| F represents the F-norm operation.

5. The radar signal modulation recognition method based on convolutional denoising and Resnet50 according to claim 1, characterized in that The residual neural network described in Step 5 is the residual neural network in the existing Resnet50 network.

6. The radar signal modulation recognition method based on convolutional denoising and Resnet50 according to claim 1, wherein The transfer learning strategy described in Step 6.2 is as follows: In the early stage of training, first train the fully connected layer in the 19-Resnet50 network, and do not change the network parameters in the residual neural network in the 19-Resnet50 network. The loaded model parameters are the parameters obtained after the 19-Resnet50 network converges during training on the ImageNet dataset. After the model converges, then train the entire deep residual neural network 19-Resnet50 to update the parameters.

Images