Construction method of radar signal identification depth migration neural network

Abstract

The invention relates to the technical field of radar signal identification, in particular to a method for constructing a radar signal identification depth migration neural network, which comprises the following steps of: S1, bispectrum analysis of a radar signal data set: by researching and analyzing bispectrum characteristics of radar signals, proposing a mature technology in image processing, and constructing a radar signal data set; a method for extracting depth migration characteristics of radar signals and using the depth migration characteristics for identifying tasks; s2, pre-training the deep convolutional neural network of the radar signal: extracting high-level semantic features from the input actually measured radar signal bispectrum image slices through forward propagation; s3, utilizing a deep convolutional neural network to extract deep migration features of bispectrum image slices of the actually measured radar signals: realizing rapid and high-precision classification and identification of the radar signals under a deep migration learning framework, and obtaining middle-layer features of the images by combining the bottom-layer features of the images through the deep convolutional neural network, and high-level features with more abstract significance are extracted by combining the middle-level features.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal recognition, and in particular relates to a method for constructing a deep migration neural network for radar signal recognition. Background Art

[0002] In the military field, radar signals are of various types and have complex waveforms. The three dominant themes of radar signal recognition in complex environments are as follows:

[0003] 1) The features designed manually based on professional knowledge cannot achieve ideal recognition results when facing radar signals of different types and under different parameter conditions, which is manifested in the poor adaptability of the features. Therefore, it is necessary to autonomously extract features with abstract meanings based on the characteristics of the signal itself in a complex environment, reflecting the autonomy of feature extraction.

[0004] 2) The same type of radar signal can present different forms in the waveform through parameter modulation, different encoding and other means. Different radar signals may present some of the same characteristics in different combat environments. The variability of radar signals makes its recognition effect maintain at a certain level for a long time, and for military needs, the recognition of radar signals requires a higher recognition accuracy. Therefore, new deep learning theories and technologies are needed to further break through the recognition accuracy of radar signals.

[0005] 3) Radar signals involve a wide measurement range and a lot of sampling data. Under certain computer hardware conditions, they have a high recognition calculation amount. However, in complex environments, opportunities are fleeting, so the radar signal recognition algorithm is required to have a faster recognition speed.

[0006] At present, some deep convolutional neural network structures have achieved the best recognition results in radar signal recognition tasks. However, the training process of the huge number of network parameters in deep convolutional neural networks is too time-consuming and is not conducive to the rapid recognition of radar signals. Summary of the invention

[0007] In view of the above problems, the present invention provides a method for constructing a deep transfer neural network for radar signal recognition, comprising the following steps:

[0008] S1. Bispectral analysis of radar signal dataset:

[0009] By studying and analyzing the dual spectrum characteristics of radar signals, a method is proposed to extract the depth migration characteristics of radar signals and use them for recognition tasks using mature image processing techniques.

[0010] S2. Pre-training of deep convolutional neural networks for radar signals:

[0011] The input measured radar signal bispectral image slices are forward propagated to extract high-level semantic features, and the network parameters are trained by back propagation of classification errors;

[0012] S3. Use deep convolutional neural network to extract the depth migration features of the bispectral image slices of the measured radar signal:

[0013] The DCNN pre-trained on the simulation data set is transferred to the task of classification and recognition of measured radar signals, and fast and high-precision classification and recognition of radar signals are achieved under the deep transfer learning framework.

[0014] Furthermore, the bispectral analysis of the radar signal data set comprises the following steps:

[0015] a. Establishment of radar signal model:

[0016] ① Based on pulse radar signals, a preliminary mathematical model of radar signals for signal sorting and identification is established;

[0017] ②Then the simulation data is based on the above signal model. In order to prevent the extracted features from being affected by the radar signal carrier frequency and noise, the individual pulse data needs to be preprocessed;

[0018] b. Bispectral analysis and estimation of radar signals:

[0019] The direct method of bispectral estimation is used to estimate, starting from a sample sequence {x(1), x(2), ..., x(N)}, and assuming that n ≤0 Or the sample sequence x(n) in the range of n≥(N+1) is always equal to zero, thereby constructing the bispectral estimation form of the random signal {x(n)};

[0020] c. Bispectral analysis of radar signals:

[0021] After converting the bispectrum into an image, the deep migration features of the radar signal are extracted using mature techniques in image processing and deep learning theory and used for recognition tasks.

[0022] Furthermore, the pre-training of the deep convolutional neural network of the radar signal comprises the following steps:

[0023] a. Structural composition and calculation of deep convolutional neural network:

[0024] Use radar simulation signal dataset to pre-train it, and migrate the pre-trained DCNN to the measured dataset to extract deep migration features about radar signals;

[0025] b. Pre-training deep convolutional neural network based on radar simulation signal dataset:

[0026] ① Before using the simulated signal data set to train DCNN, it is necessary to pre-set the objective function of the DCNN model, find the most suitable network structure parameters by minimizing the objective function, and use the chain rule to calculate the network structure parameters through the gradient descent algorithm using the error back propagation method;

[0027] ② The estimated value of the objective function calculated by the simulation signal bispectral data set.

[0028] Furthermore, the extraction of the depth migration feature of the bispectral image slice of the measured radar signal comprises the following steps:

[0029] a. Extraction of deep migration features in deep convolutional neural networks:

[0030] ① Taking AlexNet pre-trained with the simulated signal bispectral dataset as an example, the input measured radar signal bispectral is forward-calculated to obtain deep migration features from the fully connected layer before the classifier;

[0031] ②The extracted depth migration features are then directly used to train the classifier and realize the classification and recognition of the measured radar signals;

[0032] b. Fine-tuning of structural parameters in deep convolutional neural networks:

[0033] The structural parameters in the deep convolutional neural network are fine-tuned using the acquired limited bispectral image slices of the measured radar signal, and the deep migration features that are more closely related to the measured radar signal are extracted to characterize the semantic information of the measured radar signal.

[0034] Furthermore, the dual-spectrum image slice data is a long string of reflectance sequences sorted by wavelength. In addition, the reflectance data of adjacent wavelengths are correlated. Therefore, deep neural networks, especially deep neural networks for time series, can extract features from spectral data and make relevant predictions when there are labeled spectral data.

[0035] Furthermore, the deep convolutional network includes: a first convolutional layer, a second convolutional layer and a third convolutional layer, the first convolutional layer uses the first layer kernel to convolve the first layer input tensor to generate the second layer input tensor; the second convolutional layer uses the second layer kernel to convolve the second layer input tensor to generate the third layer input tensor, and the third convolutional layer similarly receives the third layer input tensor.

[0036] Further, the first layer kernel includes a first layer kernel length and a first layer kernel width;

[0037] The first layer input tensor includes the first layer input length and the first layer input width;

[0038] The second layer input tensor includes a second layer input length and a second layer input width, and the second layer kernel includes a second layer kernel length and a second layer kernel width.

[0039] Furthermore, the first convolutional layer, the second convolutional layer and the third convolutional layer are connected in parallel, and the construction layer output of the deep convolutional network is ultimately composed of the parallel construction layer outputs and the original input of the construction block in parallel; wherein the number of convolution kernel elements of the convolution blocks in the construction block are not equal to each other.

[0040] Furthermore, the radar signal classification and identification task includes a radar signal measured value evaluation information sorting module, which uses radar simulation signal samples to build a deep learning network, uses radar measured signal samples to optimize the built deep learning network, and performs measured value evaluation information feature clustering analysis on the radar simulation signal samples and the unknown radar signals to be processed.

[0041] The beneficial effects of the present invention are:

[0042] 1. The core of the present invention lies in a deep neural network structure composed of multiple layers of neural networks. This network structure can simultaneously complete the tasks of extracting features and training classifiers through the training process of network parameters. "End-to-end" feature extraction is also the main advantage of deep learning methods compared to existing feature extraction methods. Compared with other deep neural network structures, deep convolutional neural networks have their unique advantages in the process of processing visual information, and are more suitable for processing radar signal dual-spectrum image slice information. The deep convolutional neural network obtains the middle-level features of the image by combining the bottom-level features of the image, and then extracts more abstract high-level features by combining the middle-level features.

[0043] 2. The main content of the present invention is to use the radar simulation signal data set to pre-train the existing and proven effective deep convolutional neural network, obtain the corresponding network structure parameters, realize the measured radar signal bi-spectrum image slice migration feature fusion and extraction based on AlexNet, and use the obtained limited measured radar signal bi-spectrum image slices to fine-tune the structural parameters in the deep convolutional neural network, extract the deep migration features that are more closely related to the measured radar signal, and characterize the semantic information of the measured radar signal.

[0044] 3. The combination of deep learning theory and transfer learning theory can directly use the deep convolutional neural network trained in the radar simulation signal data set to extract the deep migration features of the radar measured signal (hereinafter referred to as the radar signal) under the condition of fixed network parameters. When deep learning theory is used for automatic recognition of radar signals, the deep transfer learning method is conducive to solving the following two types of problems at the same time:

[0045] First, high-precision recognition of radar signals can be achieved without sufficient radar signal data to train the massive parameters in deep convolutional neural networks.

[0046] Second, it can realize rapid recognition of radar signals without spending a lot of time training deep convolutional neural networks.

[0047] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures indicated in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0049] Figure 1 The structure diagram of each structural layer and main computing unit module of AlexNet according to an embodiment of the present invention is shown;

[0050] Figure 2 A process diagram showing the extraction of the deep migration feature of the bispectral image slice of the measured radar signal by AlexNet according to an embodiment of the present invention is shown;

[0051] Figure 3 A flow chart of a method for constructing a deep transfer neural network for radar signal recognition according to an embodiment of the present invention is shown;

[0052] Figure 4 A block diagram of a fine-tuning module for network structure parameters according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] A method for constructing a deep transfer neural network for radar signal recognition comprises the following steps:

[0055] S1. Bispectral analysis of radar signal dataset:

[0056] By studying and analyzing the dual spectrum characteristics of radar signals, a method is proposed to extract the depth migration characteristics of radar signals and use them for recognition tasks using mature image processing techniques.

[0057] S2. Pre-training of deep convolutional neural networks for radar signals:

[0058] The input measured radar signal bispectral image slices are forward propagated to extract high-level semantic features, and the network parameters are trained by back propagation of classification errors;

[0059] S3. Use deep convolutional neural network to extract the depth migration features of the bispectral image slices of the measured radar signal:

[0060] The DCNN obtained by pre-training the simulation data set is transferred to the actual radar signal classification and recognition task, and the rapid and high-precision classification and recognition of radar signals is realized under the deep transfer learning framework. Specifically, the bispectral analysis of the radar signal data set includes the following steps:

[0061] a. Establishment of radar signal model:

[0062] ① Based on pulse radar signals, a preliminary mathematical model of radar signals for signal sorting and identification is established;

[0063] ②Then the simulation data is based on the above signal model. In order to prevent the extracted features from being affected by the radar signal carrier frequency and noise, the individual pulse data needs to be preprocessed;

[0064] b. Bispectral analysis and estimation of radar signals:

[0065] The direct method of bispectral estimation is used to estimate, starting from a sample sequence {x(1), x(2), ..., x(N)}, and assuming that n ≤0 Or the sample sequence x(n) in the range of n≥(N+1) is always equal to zero, thereby constructing the bispectral estimation form of the random signal {x(n)};

[0066] c. Bispectral analysis of radar signals:

[0067] After converting the bispectrum into an image, the deep migration features of the radar signal are extracted using mature techniques in image processing and deep learning theory and used for recognition tasks.

[0068] Specifically, the pre-training of the deep convolutional neural network of the radar signal includes the following steps:

[0069] a. Structural composition and calculation of deep convolutional neural network:

[0070] Use radar simulation signal dataset to pre-train it, and migrate the pre-trained DCNN to the measured dataset to extract deep migration features about radar signals;

[0071] b. Pre-training deep convolutional neural network based on radar simulation signal dataset:

[0072] ① Before using the simulated signal data set to train DCNN, it is necessary to pre-set the objective function of the DCNN model, find the most suitable network structure parameters by minimizing the objective function, and use the chain rule to calculate the network structure parameters through the gradient descent algorithm using the error back propagation method;

[0073] ② The estimated value of the objective function calculated by the simulation signal bispectral data set.

[0074] Specifically, the extraction of the depth migration feature of the bispectral image slice of the measured radar signal includes the following steps:

[0075] a. Extraction of deep migration features in deep convolutional neural networks:

[0076] ① Taking AlexNet pre-trained with the simulated signal bispectral dataset as an example, the input measured radar signal bispectral is forward-calculated to obtain deep migration features from the fully connected layer before the classifier;

[0077] ②The extracted depth migration features are then directly used to train the classifier and realize the classification and recognition of the measured radar signals;

[0078] b. Fine-tuning of structural parameters in deep convolutional neural networks:

[0079] The structural parameters in the deep convolutional neural network are fine-tuned using the acquired limited bispectral image slices of the measured radar signal, and the deep migration features that are more closely related to the measured radar signal are extracted to characterize the semantic information of the measured radar signal.

[0080] Specifically, the dual-spectrum image slice data is a long list of reflectance sequences sorted by wavelength. In addition, the reflectance data of adjacent wavelengths are correlated. Therefore, deep neural networks, especially deep neural networks for time series, can extract features from spectral data and make relevant predictions when there are labeled spectral data.

[0081] Specifically, the deep convolutional network includes: a first convolutional layer, a second convolutional layer and a third convolutional layer, the first convolutional layer uses the first layer kernel to convolve the first layer input tensor to generate the second layer input tensor; the second convolutional layer uses the second layer kernel to convolve the second layer input tensor to generate the third layer input tensor, and the third convolutional layer similarly receives the third layer input tensor.

[0082] Specifically, the first layer kernel includes a first layer kernel length and a first layer kernel width; the first layer input tensor includes a first layer input length and a first layer input width;

[0083] The second layer input tensor includes a second layer input length and a second layer input width, and the second layer kernel includes a second layer kernel length and a second layer kernel width.

[0084] Specifically, the first convolution layer, the second convolution layer and the third convolution layer are composed in parallel, and the construction layer output of the deep convolution network is ultimately composed of the parallel construction layer output and the original input of the construction block in parallel; wherein the number of convolution kernel elements of the convolution blocks in the construction block are not equal to each other.

[0085] Specifically, the radar signal classification and recognition task includes a radar signal measured value evaluation information sorting module, which uses radar simulation signal samples to build a deep learning network, uses radar measured signal samples to optimize the built deep learning network, and performs measured value evaluation information feature clustering analysis on radar simulation signal samples and unknown radar signals to be processed.

[0086] In summary, the present application is a deep neural network structure composed of multiple layers of neural networks. This network structure can simultaneously complete the tasks of extracting features and training classifiers through the training process of network parameters. "End-to-end" feature extraction is also the main advantage of deep learning methods compared to existing feature extraction methods. Compared with other deep neural network structures, deep convolutional neural networks have their unique advantages in the process of processing visual information, and are more suitable for processing radar signal dual-spectrum image slice information. Deep convolutional neural networks obtain the middle-level features of the image by combining the underlying features of the image, and then extract more abstract high-level features by combining the middle-level features.

[0087] Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent substitutions for some of the technical features therein; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a deep transfer neural network for radar signal recognition, characterized in that: The steps include: S1. Bispectral analysis of radar signal dataset: By studying and analyzing the dual spectrum characteristics of radar signals, a method is proposed to extract the depth migration characteristics of radar signals and use them for recognition tasks using mature image processing techniques. S2. Pre-training of deep convolutional neural networks for radar signals: The input measured radar signal bispectral image slices are forward propagated to extract high-level semantic features, and the network parameters are trained by back propagation of classification errors; S3. Use deep convolutional neural network to extract the depth migration features of the bispectral image slices of the measured radar signal: The DCNN pre-trained on the simulation data set is transferred to the task of classification and recognition of measured radar signals, and fast and high-precision classification and recognition of radar signals are achieved under the deep transfer learning framework.

2. The method for constructing a deep transfer neural network for radar signal recognition according to claim 1, characterized in that: The bispectral analysis of the radar signal data set comprises the following steps: a. Establishment of radar signal model: ① Based on pulse radar signals, a preliminary mathematical model of radar signals for signal sorting and identification is established; ②Then the simulation data is based on the above signal model. In order to prevent the extracted features from being affected by the radar signal carrier frequency and noise, the individual pulse data needs to be preprocessed; b. Bispectral analysis and estimation of radar signals: The direct method of bispectral estimation is used to estimate, starting from a sample sequence {x(1), x(2), ..., x(N)}, and assuming that n ≤0 Or the sample sequence x(n) in the range of n≥(N+1) is always equal to zero, thereby constructing the bispectral estimation form of the random signal {x(n)}; c. Bispectral analysis of radar signals: After converting the bispectrum into an image, the deep migration features of the radar signal are extracted using mature techniques in image processing and deep learning theory and used for recognition tasks.

3. The method for constructing a deep transfer neural network for radar signal recognition according to claim 1, characterized in that: The pre-training of the deep convolutional neural network of the radar signal comprises the following steps: a. Structural composition and calculation of deep convolutional neural network: Use radar simulation signal dataset to pre-train it, and migrate the pre-trained DCNN to the measured dataset to extract deep migration features about radar signals; b. Pre-training deep convolutional neural network based on radar simulation signal dataset: ① Before using the simulated signal data set to train DCNN, it is necessary to pre-set the objective function of the DCNN model, find the most suitable network structure parameters by minimizing the objective function, and use the chain rule to calculate the network structure parameters through the gradient descent algorithm using the error back propagation method; ② The estimated value of the objective function calculated by the simulation signal bispectral data set.

4. The method for constructing a deep transfer neural network for radar signal recognition according to claim 1, characterized in that: The method of extracting the depth migration feature of the bispectral image slice of the measured radar signal comprises the following steps: a. Extraction of deep migration features in deep convolutional neural networks: ① Taking AlexNet pre-trained with the simulated signal bispectral dataset as an example, the input measured radar signal bispectral is forward-calculated to obtain deep migration features from the fully connected layer before the classifier; ②The extracted depth migration features are then directly used to train the classifier and realize the classification and recognition of the measured radar signals; b. Fine-tuning of structural parameters in deep convolutional neural networks: The structural parameters in the deep convolutional neural network are fine-tuned using the acquired limited bispectral image slices of the measured radar signal, and the deep migration features that are more closely related to the measured radar signal are extracted to characterize the semantic information of the measured radar signal.

5. The method for constructing a deep transfer neural network for radar signal recognition according to claim 4, characterized in that: The dual-spectrum image slice data is a long sequence of reflectance sequences sorted by wavelength. In addition, the reflectance data of adjacent wavelengths are correlated. Therefore, deep neural networks, especially deep neural networks for time series, can extract features from spectral data and make relevant predictions when there are labeled spectral data.

6. The method for constructing a deep transfer neural network for radar signal recognition according to claim 1, characterized in that: The deep convolutional network includes: a first convolutional layer, a second convolutional layer and a third convolutional layer. The first convolutional layer uses the first layer kernel to convolve the first layer input tensor to generate the second layer input tensor; the second convolutional layer uses the second layer kernel to convolve the second layer input tensor to generate the third layer input tensor, and the third convolutional layer similarly receives the third layer input tensor.

7. The method for constructing a deep transfer neural network for radar signal recognition according to claim 6, characterized in that: The first layer kernel includes a first layer kernel length and a first layer kernel width; the first layer input tensor includes a first layer input length and a first layer input width; The second layer input tensor includes a second layer input length and a second layer input width, and the second layer kernel includes a second layer kernel length and a second layer kernel width.

8. The method for constructing a deep transfer neural network for radar signal recognition according to claim 6, characterized in that: The first convolutional layer, the second convolutional layer and the third convolutional layer are connected in parallel, and the construction layer output of the deep convolutional network is ultimately composed of the parallel construction layer outputs and the original input of the construction block in parallel; wherein the number of convolution kernel elements of the convolution blocks in the construction block are not equal to each other.

9. The method for constructing a deep transfer neural network for radar signal recognition according to claim 1, characterized in that: The radar signal classification and recognition task includes a radar signal measured value evaluation information sorting module, which uses radar simulation signal samples to build a deep learning network, uses radar measured signal samples to optimize the built deep learning network, and performs measured value evaluation information feature clustering analysis on radar simulation signal samples and unknown radar signals to be processed.

Images