Radar emitter signal recognition method based on CNN-BLS network

Abstract

A kind of radar radiation source signal identification method based on CNN-BLS network, including constructing data set, pre-processing to time-frequency image, constructing improved CNN-BLS network, training CNN-BLS network and identifying radar radiation source signal;The present application effectively fuses convolutional neural network (CNN) and width learning network (BLS), realizes the classification of radar radiation source signal identification, effectively balances the calculation cost and recognition accuracy of neural network model, with fast training speed, high recognition accuracy and other characteristics.In reducing the training time of network model, improve the real-time, improve the accuracy of radar radiation source signal identification classification, ensure the accuracy of radar radiation source signal identification classification, realize the real-time and accuracy of radar radiation source signal identification classification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electronic countermeasure, and particularly relates to a radar emitter signal recognition method based on a CNN-BLS network. The present application is used for recognizing and classifying radar emitter signals in radar emitter recognition, and ensures the accuracy and real-time performance of radar emitter signal recognition and classification. BACKGROUND

[0002] Radar emitter recognition, as an important part of electronic countermeasure reconnaissance, plays an important role in radar target type judgment and threat level alarm. Radar signal intra-pulse modulation type recognition is an important part of radar reconnaissance system, and accurate recognition of signal type is helpful for judging enemy radar type, threat level and function. In the modern increasingly dense and complex electromagnetic environment, new radar systems emerge in an endless stream, and signal styles are complex and changeable. Radar emitter signal recognition is facing severe challenges. The traditional radar emitter signal recognition method based on five pulse description words and other characteristic parameters cannot meet the demand of quickly and accurately distinguishing radar emitter signals on the battlefield. The method of manually designed feature extraction needs to rely on professional knowledge and experience, and has defects such as strong pertinence and failure under low signal-to-noise ratio. How to automatically obtain deep essential features conducive to signal classification from data has become a research hotspot of radar signal recognition.

[0003] Deep learning automatically obtains more classification features by training massive data and constructing learning models. In recent years, domestic and foreign scholars have applied deep learning technology to radar emitter signal recognition. Under the condition of low signal-to-noise ratio, this method is more robust than the recognition method of artificial rule construction features. Ali et al. used unsupervised feature learning and fully connected neural networks for intra-pulse modulation recognition, which achieved a recognition rate of 94.5% under the condition of-5~15 dB signal-to-noise ratio (see Ali A, Fan Y. Unsupervised feature learning and automatic modulation classification using deep learning model[J]. Physical Communication, 2017, 25(PT.1): 75-84.). Although this method has greatly improved the recognition rate compared with traditional radar emitter signal recognition methods, it can automatically extract the features required for signal classification, but due to the small number of neural network layers, it cannot fully extract data features, which may cause feature learning overfitting or underfitting. And because it is unsupervised learning, it is easily disturbed by data samples, and it is too sensitive to small fluctuations in training data, which may cause the prediction performance of the network to decline. Zhou et al. designed an automatic radar signal recognition method based on a restricted Boltzmann machine to extract radar feature parameters and recognize radar signals (see Zhou D, Wang X, Tian Y, et al. A novel radar signal recognition method based on a deep restricted Boltzmann machine[J]. Engineering Review, 2017, 37(2): 165-171). Subsequently, Jing Bojun proposed obtaining the time-frequency image of the signal, and automatically learning the features of the signal time-frequency image through a convolutional neural network (CNN) to classify and recognize (see Radar Emitter Recognition Technology Based on Deep Learning[D]. Xi'an University of Electronic Science and Technology, 2017). Meng et al. constructed an "End to End" convolutional neural network modulation signal recognition method, and used two-stage training method and transfer learning idea to improve the training efficiency.Finally, higher accuracy and robustness are obtained, and the computational overhead is much smaller than that of traditional methods (see: Meng F, Chen P, Wu L, et al. Automatic Modulation Classification: A Deep Learning Enabled Approach [J]. IEEE Transactions on Vehicular Technology, 2018, 67(11): 10760-10772). Experiments show that the convolutional neural network has stronger feature extraction capability and can obtain higher signal recognition accuracy and stronger anti-interference ability. Since then, many people have proposed radar emitter signal recognition methods based on improved convolutional neural networks, such as Qin Xin et al. who use time-frequency analysis to convert signals into time-frequency feature images, then use the dilated residual network to identify these two-dimensional time-frequency feature maps, and get very effective recognition results.

[0004] Although the radar emitter signal recognition method based on deep convolutional network has made certain breakthroughs in recognition performance, due to the deep network depth of the deep learning model, it needs more time to train the network model, resulting in high computational cost and increasing hardware overhead cost. In addition, in the battlefield, high computational cost and high-cost hardware overhead will also affect the rapid response ability of electronic countermeasures and the real-time countermeasure ability of high precision.

[0005] The Broad Learning System (BLS) is first proposed by Professor Chen Junlong of the University of Macau and his students in 2018, aiming to solve the problems of complex structure, huge parameter quantity and long calculation time existing in most deep learning networks (for details, see: Chen C, Liu Z. Broad Learning System: An Effective and Efficient Incremental Learning System Without the Need for Deep Architecture [J]. IEEE Transactions on Neural Networks & Learning Systems, 2018, 29 (99): 10-24.). The BLS method has been applied to many fields such as image classification, sequence prediction, etc. However, although the BLS method has faster calculation speed in image recognition and can obtain classification results in very short calculation time, it can effectively alleviate the problem of time consumption of training model in deep learning network, but its performance in high-dimensional data is not satisfactory, and it cannot effectively extract location-related information and other detailed features from image pixels, and the accuracy in radar emitter signal recognition classification needs to be improved.

[0006] In summary, although the existing radar emitter signal recognition method based on deep convolutional network has made certain breakthroughs in recognition performance, these network models have high complexity, huge parameter quantity, large hardware overhead cost, and problems of long model training time and excessive parameter quantity. How to effectively balance the calculation cost and recognition accuracy of neural network model has become the key to whether the radar emitter signal recognition method can be applied in engineering. The existing BLS method can effectively alleviate the problem of long model training time in deep neural network, but its feature learning ability in high-dimensional data is not satisfactory, and the accuracy of recognition classification is not high. SUMMARY

[0007] The purpose of the present application is to propose a radar emitter signal recognition method based on CNN-BLS network. For radar emitter signal recognition classification, a radar emitter signal recognition method based on CNN-BLS network is first proposed, which effectively fuses convolutional neural network (CNN) and broad learning network (BLS), realizes radar emitter signal recognition, effectively balances the calculation cost and recognition accuracy of neural network model, improves the accuracy of radar emitter signal recognition classification, reduces the training time of network model, and ensures the real-time and accuracy of radar emitter signal recognition classification.

[0008] The present application is realized by the following technical solutions.

[0009] The radar radiation source signal recognition method based on the CNN-BLS network comprises the following steps:

[0010] Step 1, constructing a data set, performing Choi-Williams time-frequency analysis on the radar radiation source signal received by the intercepting device, so as to convert the radar signal into a two-dimensional time-frequency image.

[0011] Step 2, preprocessing the time-frequency image, first performing gray processing on the radar signal time-frequency image to reduce image redundancy information, then using an image opening operation method to perform noise reduction processing on the image, and finally performing normalization processing to adjust the size of the time-frequency image. Obtain the preprocessed time-frequency image set, and divide the preprocessed time-frequency image set into a training set, a validation set and a test set according to a preset proportion.

[0012] Step 3, constructing an improved CNN-BLS network. The CNN-BLS network is composed of two parts, which are divided into a feature extraction module and a classification recognition module. A designed CNN network is used to extract as much detailed feature information as possible, and then the extracted features are sent to a width learning network (Broad Learning System, BLS) to quickly obtain a recognition classification result.

[0013] Step 4, training the CNN-BLS network with the training set samples in step 2.

[0014] Step 5, recognizing the radar radiation source signal. The test set data in step 2 is input into the CNN-BLS network trained in step 4, the classification and recognition of the radar radiation source signal are completed, the recognition classification result is obtained, and the recognition accuracy is calculated.

[0015] The improved CNN-BLS network constructed in step 3 comprises a feature extraction module and a classification recognition module.

[0016] The feature extraction module network consists of multiple convolutional layers using depthwise separable convolution (depsconv). Each layer uses a stride (the step size of the convolution operation) for downsampling to reduce data dimensionality. To avoid internal covariate shifts in the activation input value distribution during training, which could lead to vanishing gradients at the lower layers during backpropagation, the activation input values ​​are first fed into a batch normalization (BN) layer before each non-linear transformation. In the later layers of the network model, large convolutional kernels are used instead of small ones. Using large kernels during downsampling allows for a larger receptive field, enabling the network to extract more detailed features. The kernel size is N×N, where N can be a positive integer (3, 5, 7, ...); the number of convolutional layers can also be a positive integer (5, 6, 7, 8, 9, ...).

[0017] After the feature extraction module fully extracts the detailed features from the image, the extracted features are fed into the classification and recognition module. The classification and recognition module consists of a width learning network. The features extracted by the feature extraction module are fed into the width learning network as the input matrix. The input data is transformed into feature nodes through linear mapping, and then the feature nodes are transformed into enhancement nodes through non-linear mapping. Finally, the classification result is obtained by outputting weights through the width learning network. The specific classification and recognition process is as follows:

[0018] (a) Feature node generation. The linear mapping of feature nodes is shown in equation (1):

[0019]

[0020] Where X is the input matrix, which is processed by a linear function. The mapping produces the i-th set of mapping features Z i (i = 1, 2, 3, ..., n), These are random weight coefficients; all n feature nodes in a feature node group can be represented as Z. n =[Z1,Z2…Z n ].

[0021] (b) Enhancement node generation. Enhancement nodes, denoted as H, are obtained by mapping feature nodes through a nonlinear function. j (j=1,2,...,m) is shown in equation (2):

[0022]

[0023] Where j represents the number of groups that generate augmented nodes, and the total number of m groups of augmented nodes can be represented as H. m = [H1,H2,…..,H m ].

[0024] (c) obtaining the weights of the width learning network. The outputs of the feature nodes and the enhanced nodes are connected, and the total input of the output layer is denoted as A, as shown in formula (3):

[0025] Y=[Z n ,H m ]W m =AW (3)

[0026] Y is the output layer, W m is the weight connection between the output layer and the feature nodes and the enhanced nodes. The problem of obtaining the weights of the width learning network is converted into a linear equation system problem, and the weights W are obtained by calculating the pseudo-inverse. Considering that the input matrix of the network is not necessarily invertible, the pseudo-inverse of the weights W of the BLS network is obtained by the ridge regression pseudo-inverse method, and the pseudo-inverse of A can be obtained by formula (4):

[0027]

[0028] Therefore, the output connection weights W of the network can be obtained by formula (5), and the output result is obtained:

[0029] W=A + Y (5)

[0030] The CNN-BLS network is trained according to the step 4. The training includes:

[0031] The training set sample of the preprocessed time-frequency graph in step 2 is input into the CNN-BLS network to train the network. An optimization algorithm is used to update the network weight value, for example, the Adam optimization algorithm and the SGD optimization algorithm, a cross-entropy loss function is used, the maximum training round Q is set to a positive integer greater than or equal to 60, a learning rate decay strategy is added, the learning rate decay period L is any positive integer in the range of the maximum training round, and the learning rate is reduced to half of the current learning rate after every L rounds of training. The recognition accuracy of the verification set is used as the standard, and the network model with the highest recognition accuracy is saved. The network hyperparameters are updated by continuous iteration, and when the number of iterations reaches the set number of iterations, the training process of the neural network is completed.

[0032] The radar radiation source signal recognition method based on the CNN-BLS network provided by the application effectively balances the calculation cost and recognition accuracy of the neural network model, has the characteristics of fast training speed and high recognition accuracy. The radar radiation source signal recognition method based on the CNN-BLS algorithm reduces the model training time and improves the real-time performance, ensures the accuracy of the radar radiation source signal recognition classification, and realizes the real-time and accuracy of the radar radiation source signal recognition classification. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1A CNN-BLS network model structure diagram is shown in the figure, X is a radar emitter signal time-frequency diagram, the dashed box is a feature extraction module, and Y is an output result.

[0034] Figure 2 A depth separable convolution principle diagram is shown in the figure.

[0035] Figure 3 A feature extraction module structure diagram is shown in the figure, Input image size is an input image size, depsconv is a depth separable convolution, stride is a convolution operation step length, and conv_n (n = 1-7) is an nth convolution layer. DETAILED DESCRIPTION

[0036] To make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0037] As a feasible embodiment of the present application, radar sample signals of different intra-pulse modulation modes in equal number are generated every 2 dB under a signal-to-noise ratio of -10 dB to 8 dB, and the radar sample signals of different intra-pulse modulation modes include conventional pulse signals, linear frequency modulation signals, frequency encoding signals, phase encoding signals and hybrid modulation signals.

[0038] The present application takes 10 radar intra-pulse modulation signals as embodiments for further description, including single-frequency rectangular pulse signals CW, linear frequency modulation signals LFM, nonlinear frequency modulation signals NLFM, binary phase encoding signals BPSK, multi-phase encoding signals Frank, P1, P2, P3, P4 and frequency encoding signals Costas signals. In a range of -10 dB to 8 dB, 10 signal-to-noise ratios are equally spaced at intervals of 2 dB, 100 modulation signals are collected as original signals for each of the 10 types of radar intra-pulse modulation signals under each signal-to-noise ratio, and a total of 10,000 random radar intra-pulse modulation signals are obtained. For each type of signal modulation mode, a range randomization parameter is used, Choi-Williams time-frequency analysis is performed on the random radar intra-pulse modulation signals to obtain a time-frequency image, normalization and grayscale processing are performed on the time-frequency image to obtain a corresponding time-frequency domain grayscale image, and the obtained 10,000 grayscale time-frequency images are used as radar intra-pulse modulation signal data sets for training and testing the CNN-BLS network model.

[0039] Step 1, constructing a data set, performing Choi-Williams time-frequency analysis on the radar radiation source signal received by the intercepting device, converting the radar signal into a two-dimensional time-frequency image.

[0040] Step 2, preprocessing the time-frequency image, first performing gray scale processing on the radar signal time-frequency image to reduce image redundancy information, then using image opening operation method to denoise the image, and finally performing normalization processing to adjust the size of the time-frequency image. The preprocessed time-frequency image set is divided into training set, validation set and test set according to the preset proportion.

[0041] Step 3, constructing an improved CNN-BLS network. The CNN-BLS network is composed of two parts, which are feature extraction module and classification identification module. The designed CNN network is used to extract as much detailed feature information as possible, and then the extracted features are sent to the width learning network BLS to quickly obtain the recognition classification result.

[0042] Step 4, training the CNN-BLS network with the training set samples in step 2.

[0043] Step 5, identifying the radar radiation source signal. The test set data in step 2 is input into the CNN-BLS network trained in step 4 to complete the classification and identification of the radar radiation source signal, obtain the recognition classification result and calculate the recognition accuracy.

[0044] The step 3 constructs an improved CNN-BLS network, which includes a feature extraction module and a classification identification module.

[0045] The feature extraction module network is composed of 7 convolutional layers, and the convolutional layers use depthwise separable convolution (depsconv). Each layer uses stride (convolution operation step length) for down sampling to reduce the dimension of the data. In order to avoid the internal covariate shift of the activation input value distribution of each layer during the training process, which leads to the disappearance of the gradient of the bottom layer of the neural network during back propagation, the activation input value is input into the BN layer before nonlinear change, and the batch normalization operation is performed on the activation input value. Large convolution kernels are used in the middle and later layers of the network to replace small convolution kernels. When down sampling, large convolution kernels are used to obtain a larger receptive field while allowing the network to extract more detailed features. The first, fourth and fifth convolutional layers use convolution kernels with a size of 3x3 and a stride of 1, and the second, third, sixth and seventh convolutional layers use convolution kernels with a size of 5x5 and a stride of 2. The input image size of the feature extraction model is 128x128, and the channel number of each convolutional layer is [16, 32, 64, 64, 64, 128, 256].

[0046] After the feature extraction module fully extracts the detailed features from the image, the extracted features are fed into the classification and recognition module. The classification and recognition module consists of a width learning network. The features extracted by the feature extraction module are fed into the width learning network as the input matrix. The input data is transformed into feature nodes through linear mapping, and then the feature nodes are transformed into enhancement nodes through non-linear mapping. Finally, the classification result is obtained by outputting weights through the width learning network. In this embodiment, the number of feature node groups n is set to 9, with 18 feature nodes in each group, and the total number of enhancement nodes is 2000. The specific classification and recognition process is as follows:

[0047] (a) Feature node generation. The linear mapping of the feature layer is shown in equation (1):

[0048]

[0049] Where X is the input matrix, which is processed by a linear function. The mapping produces the i-th set of mapping features Z i (i = 1, 2, 3, ..., n), These are random weight coefficients; all n feature nodes in a feature node group can be represented as Z. n =[Z1,Z2…Z n ].

[0050] (b) Enhancement node generation. Enhancement nodes, denoted as H, are obtained by mapping feature nodes through a nonlinear function. j (j=1,2,...,m) is shown in equation (2):

[0051]

[0052] Where j represents the number of groups that generate augmented nodes, then all m groups of augmented nodes can be represented as H. m =[H1,H2,…..,H m ].

[0053] (c) Obtain the weights of the width learning network. After concatenating the outputs of the feature mapping layer and the enhancement layer, we can consider it as the total input of the output layer, denoted as A. As shown in equation (3):

[0054] Y = [Z] n H m W m =AW (3)

[0055] Y is the output layer, W mis the weight connecting the output layer and the feature mapping node and the enhanced node. The problem of calculating the weight of the width learning network is converted into a linear equation problem, and the weight W is calculated by the pseudo-inverse method. Considering that the input matrix of the network is not necessarily invertible, the BLS network weight is obtained by the pseudo-inverse method of ridge regression, and the pseudo-inverse of A can be obtained by formula (4):

[0056]

[0057] Therefore, the output connection weight W of the network can be calculated by formula (5), and the output result is obtained.

[0058] W=A + Y (5)

[0059] The step 4 trains the radar radiation source signal recognition network. It includes:

[0060] The training set sample of the preprocessed time-frequency graph in step 2 is input into the radar radiation source recognition network to train the network. The Adam optimization algorithm is used to update the network weight, the cross-entropy loss function is used, and the maximum training number is set to 100 rounds. Learning rate decay strategy is added, and the learning rate decay period is 40. After every 40 rounds of training, the learning rate is reduced to half of the current learning rate. The identification accuracy of the verification set is used as the standard, and the network model with the highest identification accuracy is saved. Through continuous iteration and update of network hyperparameters, when the iteration number reaches the set iteration number, the training process of the neural network is completed.

[0061] The radar radiation source signal recognition classification accuracy and model training time results obtained by the embodiment are as follows. In order to further reflect the advantages of the present application, the method of the present application is compared with VGG16, ResNet18, DRN and other traditional deep convolutional neural network algorithms. The overall classification recognition accuracy of the model under the same experimental conditions and the time consumption required for model training are compared. The results of Table 1 and Table 2 show that the radar radiation source signal recognition based on the CNN-BLS algorithm provides the highest classification recognition accuracy and the least model training time consumption.

[0062] Table 1 Radar radiation source signal recognition accuracy under different algorithms

[0063]

[0064] Table 2 Model training time consumption (100epoch) under different algorithms

[0065]

[0066] The above merely expresses the preferred embodiments of the present application, which are described in more detail and in a more specific manner, but should not be understood as limiting the scope of the patent of the present application. It should be noted that, for those of ordinary skill in the art, several modifications, improvements and substitutions can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1.A radar emitter signal recognition method based on a CNN-BLS network, characterized by The method comprises the following steps: Step 1, constructing a data set: performing Choi-Williams time-frequency analysis on the radar radiation source signals received by the intercepting device, so as to convert the radar signals into two-dimensional time-frequency images; Step 2, preprocessing the time-frequency images: first, performing gray-scale processing on the radar signal time-frequency images to reduce image redundancy information, then performing noise reduction processing on the images using an image opening operation method, and finally performing normalization processing to adjust the size of the time-frequency images; obtaining the preprocessed time-frequency image set, and dividing the preprocessed time-frequency image set into a training set, a validation set and a test set according to a preset proportion; Step 3, constructing an improved CNN-BLS network: the CNN-BLS network is composed of two parts, namely a feature extraction module and a classification identification module, a designed CNN network is used to extract more detailed feature information, and then the extracted features are sent to a width learning network BLS to quickly obtain a recognition classification result; Step 4, training the CNN-BLS network using the training set samples in step 2; Step 5, identifying the radar radiation source signals: inputting the test set data in step 2 into the CNN-BLS network trained in step 4, completing the classification and identification of the radar radiation source signals, obtaining the classification and identification results and calculating the identification accuracy; The improved CNN-BLS network constructed in step 3, Comprises a feature extraction module and a classification identification module; The feature extraction module network is composed of multiple convolutional layers, the convolutional layers use depth separable convolution, and stride is used between each layer for down-sampling to reduce the dimension of the data; In order to avoid internal covariate shift of the activation input value distribution of each layer during the training process, causing the gradient of the neural network bottom layer to disappear during back propagation, the activation input value is input into the BN layer before nonlinear change in each layer, and the batch normalization operation is performed on the activation input value; large convolution kernels are used in the middle and later layers of the network to replace small convolution kernels, and large convolution kernels are used for down-sampling to obtain a larger receptive field while allowing the network to extract more detailed features; the size of the convolution kernel is N*N, and N is a positive integer 3, 5, 7, …; the number of convolutional layers is a positive integer 5, 6, 7, 8, 9, …; After the feature extraction module is used to fully extract the detailed features in the image, the extracted features are sent to the classification identification module; the classification identification module is composed of a width learning network, the features extracted by the feature extraction module are sent to the width learning network as an input matrix, the input data is linearly mapped into feature nodes, then the feature nodes are nonlinearly mapped into enhanced nodes, and finally the classification result is obtained through the output weight of the width learning network; the classification identification process is as follows: (a) Generation of feature nodes: the linear mapping of the feature nodes is shown in formula (1): wherein X is an input matrix, the input matrix is subjected to a linear function The mapping generates the i-th group of mapping features Z i (i = 1, 2, 3…n), is a random weight coefficient, and all n groups of feature nodes can be expressed as Z n = [Z1, Z2…Z n ]; (b) enhanced node generation: the feature nodes are mapped by a nonlinear function to obtain enhanced nodes, denoted as H j (j = 1, 2,..., m) as shown in equation (2): where j represents the number of groups of enhanced nodes generated, and all m groups of enhanced nodes can be represented as H m = [H1, H2,...., H m ] ; (c) Obtaining the weight of the width learning network: connecting the outputs of the feature nodes and the enhanced nodes, setting it as the total input of the output layer, denoted as A, as shown in formula (3): Y = [Z n ,H m ]W m = AW (3) Y is the output layer, W m is the weight connecting the output layer and the feature nodes and the enhancement nodes; the weight W is calculated by solving the pseudo-inverse method to convert the weight problem of the BLS network into a linear equation set problem; considering that the input matrix of the network is not necessarily invertible, the weight of the BLS network is obtained by the pseudo-inverse method of ridge regression, and the pseudo-inverse of A can be obtained by formula (4): Therefore, the output connection weight W of the network can be obtained through formula (5), and then the output result is obtained: W = A + Y (5) The training CNN-BLS network process described in step 4 comprises: inputting the training set samples of the preprocessed time-frequency diagram in step 2 into the CNN-BLS network to train the network; updating the network weight values by using an optimization algorithm, using a cross-entropy loss function, and setting the maximum number of training rounds Q as a positive integer greater than or equal to 60; adding a learning rate decay strategy, and setting the learning rate decay period L as any positive integer within the range of the maximum number of training rounds, and the learning rate is reduced to half of the current learning rate every L rounds of training; taking the verification set recognition accuracy as the standard, and saving the network model with the highest recognition accuracy; and completing the neural network training process by continuously updating the network hyperparameters and when the number of iterations reaches the set number of iterations.

Images