Electromagnetic signal recognition method based on one-dimensional complex-valued residual network

Abstract

The application provides an electromagnetic signal recognition method based on a one-dimensional complex residual network, converts a communication signal received by a receiver into signal data in an int16 format, performs Fourier transform, draws a frequency spectrum diagram of the signal, estimates a carrier frequency of the signal, resamples the signal, performs intermediate frequency filtering, and then performs power normalization, trains a data set after power normalization processing through a one-dimensional complex residual network, and calculates the accuracy of the model on a test set, thereby solving the problem that the feature of a signal is lost due to training of a neural network by using real number representation of a communication signal in the prior art, and the residual network can avoid the gradient vanishing problem, so that a deeper network can be designed to extract the original feature of the communication signal, and a higher accuracy of modulation recognition is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of communication, in particular to a modulation recognition method of wireless communication signals, which is suitable for identifying the modulation mode of electromagnetic signals by using deep learning technology. BACKGROUND

[0002] Automatic modulation recognition (AMR) of electromagnetic signals, as an intermediate process between signal detection and signal demodulation, provides signal modulation information, and plays a key role in civilian and military applications such as cognitive radio, signal recognition, threat assessment and spectrum monitoring.

[0003] Traditional electromagnetic signal recognition methods can be divided into two categories: likelihood-based methods and feature-based methods. Likelihood-based methods usually obtain higher accuracy and minimize the probability of error, but such methods have high delay classification or require complete prior knowledge, and the amount of calculation is large. Traditional feature-based methods mostly use decision trees, support vector machines and other classifiers, which have low classification accuracy.

[0004] In recent years, artificial intelligence methods represented by deep learning use the learning ability of the network, do not need to extract expert features, and have made significant progress in electromagnetic signal recognition. However, current deep learning methods mostly train neural networks using real number representations of signals, and less use the real and imaginary part correlation information of complex numbers, thereby limiting the further improvement of electromagnetic signal recognition accuracy. SUMMARY

[0005] In order to overcome the shortcomings of the prior art, the present application provides an electromagnetic signal recognition method based on one-dimensional complex residual network. In view of the defects of the prior art, the purpose of the present application is to provide a communication signal modulation recognition method based on one-dimensional complex residual neural network (ResNet). The problem of feature loss caused by training neural networks using real number representations of communication signals in existing research is solved. At the same time, the residual network can avoid the problem of gradient disappearance, so that a deeper network can be designed to extract the original features of the communication signal, and a higher accuracy of modulation recognition is achieved.

[0006] The technical solution adopted by the present application to solve its technical problems comprises the following steps:

[0007] Step 1: converting the communication signal received by the receiver into int16 format signal data;

[0008] Step 2: performing Fourier transform on the signal data of step 1, drawing the frequency spectrum of the signal, and estimating the carrier frequency of the signal;

[0009] Step 3: resample the signal data generated in step 1 with four times the carrier frequency;

[0010] Step 4: perform intermediate frequency filtering on the signal obtained in step 3 according to the carrier frequency estimated in step 2;

[0011] Step 5: perform power normalization processing on the signal filtered in step 4;

[0012] Step 6: divide the data set after power normalization processing into a training set and a test set, design a one-dimensional complex-valued residual network, use the one-dimensional complex-valued residual network for training, then use the test set to input into the trained deep complex-valued attention mechanism neural network for testing, and obtain the corresponding digital signal modulation mode;

[0013] Step 7: model testing

[0014] The divided test set is input into the trained model, and the output class of the model is compared with the label class of each data to judge the accuracy of the model on the test set.

[0015] The one-dimensional complex-valued residual network in step 6 is composed of six sub-modules, which are a first one-dimensional complex-valued convolution module, a first C-ResNet Block module, a second C-ResNet Block module, a third C-ResNet Block module, a fourth C-ResNet Block module, and a full connection layer module, and the six sub-modules are connected in series.

[0016] The first C-ResNet Block module includes three complex-valued ResNet building blocks, the second C-ResNet Block module includes four complex-valued ResNet building blocks, the third C-ResNet Block module includes six complex-valued ResNet building blocks, and the fourth C-ResNet Block module includes three complex-valued ResNet building blocks.

[0017] The complex-valued ResNet building block includes a first one-dimensional complex-valued convolution layer, a first one-dimensional complex-valued Batchnorm layer, a first complex-valued ReLu activation function, a second one-dimensional complex-valued convolution layer, a second one-dimensional complex-valued Batchnorm layer, a first complex-valued downsample, and a second complex-valued ReLu activation function, and the above complex-valued convolution layers and activation functions are connected in series.

[0018] In the complex-valued convolution layer, in order to use real-valued algorithms to simulate complex-valued algorithms internally, a complex-valued convolution kernel weight matrix W=A+iB and a complex-valued vector s=x+iy are defined, which represent I / Q signals respectively, and the obtained convolution kernel is used to convolve the complex vector:

[0019]

[0020] where A and B are the real and imaginary parts of the complex convolution kernel weight matrix respectively, and x and y are the real and imaginary parts of the complex vector respectively;

[0021] Then, the real and imaginary parts of the convolution result are obtained as follows:

[0022]

[0023]

[0024] Each convolution layer is followed by a complex batch normalization layer, and batch normalization is an important technique for optimizing data models. The real-valued complex batch normalization layer is represented as:

[0025]

[0026] where, and V are the mean and variance of a Batch data respectively; γ and β are trainable parameters, and σ is a very small value to prevent the denominator from being zero;

[0027] However, for a one-dimensional complex-valued residual network:

[0028]

[0029] The learnable shift parameter β and the scale parameter γ of the complex batch normalization are set, and the scale parameter γ is given by:

[0030]

[0031] The shift parameter β in the shift parameter γ and the scale parameter γ ri will be initialized to zero, and γ in the scale parameter γ rr and γ ii are initialized to

[0032] In step 6, before training the one-dimensional complex-valued residual network, weight initialization is performed, and correct initialization of the one-dimensional complex-valued residual network is crucial to reduce the risk of gradient vanishing or explosion. The complex weight is represented as:

[0033] W = Real{W} + iImag{W}

[0034] When W is symmetrically distributed around 0, the variance of W is estimated from the parameter σ of the Rayleigh distribution, and the weight amplitude is initialized according to the Rayleigh distribution, with an expected parameter of 0 and a variance parameter of 2σ 2 , and the parameter σ will be set differently according to different neural network architectures.

[0035] In the one-dimensional complex-valued residual network, the deeper neural network can extract more discriminative features for automatic modulation classification, but as the neural network deepens, the training accuracy will gradually saturate and deteriorate; the ResNet network can effectively solve the gradient disappearance problem through its unique residual structure, so it is designed as a deeper network. The ResNet building block contains two mappings, residual mapping and identity mapping, where the residual mapping is described as:

[0036] F(z) = H(z) - z

[0037] Where represents the expected underlying mapping of the ResNet block, under this structure, the network only needs to learn the difference between the input and the output, which simplifies the learning difficulty and solves the problem of deep gradient disappearance.

[0038] The one-dimensional complex-valued residual network adopts C-ResNet18 or C-ResNet34, and the structure is as shown in Figure 1 (a) and Figure 1 (b) shown.

[0039] The loss function of the one-dimensional complex-valued residual network adopts the cross-entropy loss function, the optimization function selects Adam, the initial learning rate lr=0.001, and the learning rate modulation strategy adopts the equal interval adjustment learning rate StepLR, the adjustment interval is 50Epochs, and the adjustment multiple is 0.1.

[0040] The beneficial effects of the present application are as follows: first, the communication signal is represented by a complex number, and the real part and the imaginary part of the signal are mutually dependent under any phase change caused by the time shift effect. The real-valued network model will consider that the real part and the imaginary part of the signal are independent, while the complex-valued neural network model will consider the correlation between the real part and the imaginary part of the signal. Second, the amplitude and the phase are important to the learning goal, so it is meaningful to use the complex-valued neural network model. The complex signal represents a single variable complex parameter instead of two double variable real parameters, so we have a less free degree solution space. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The one-dimensional complex-valued residual network structure proposed by the present application, Figure 1 (a) is the C-ResNet18 network structure, Figure 1 (b) is the C-ResNet34 network structure.

[0042] Figure 2 The effect diagram of comparing the one-dimensional complex-valued residual network with the corresponding real-valued network based on the RadioML2016.10B data set. DETAILED DESCRIPTION

[0043] The present application will be further described below in conjunction with the drawings and examples.

[0044] As shown in Figure 1 A method for electromagnetic signal recognition based on one-dimensional complex residual network. The experiment of the application uses the benchmark open source dataset RML2016.10B. The signal-to-noise ratio range of the three datasets is from -20dB to 18dB, and the signal is collected at a sampling rate of 4 samples per symbol. Each data sample is an IQ sequence with 128 sampling points. In the generation process, a large number of real speech signals are used, and various channel effects such as carrier frequency offset (CFO), phase offset, Gaussian white noise, fading and multipath are simulated by using the dynamic channel model in GNU Radio.

[0045] Step 1: Sampling the public standard modulation signal dataset RML2016.10B dataset.

[0046] Step 2: Generation of modulation signal time-frequency spectrum data set;

[0047] The RML2016.10B dataset simulates various modulation signals in real life by using the software radio software GUN Radio, including 11 modulation signals: 8PSK, BPSK, AM-DSB, QPSK, QAM16, QAM64, CPFSK, GFSK, 4PAM and WBFM. The signal-to-noise ratio coverage is from -20dB to +18dB, with an interval of 2dB, and the sampling length is 128. All signal data are collected from the signal at a sampling rate of 4 samples per symbol and 1M / s. The received signal is affected by the defects of the wireless channel. Each signal data is saved as a 2x128 matrix, and the two rows of data correspond to the in-phase part and the quadrature part of the complex signal sample.

[0048] Step 2: Data annotation;

[0049] The dataset contains 10 modulation methods, and each signal data is labeled with SNR and modulation method.

[0050] Step 3: Dataset division;

[0051] In order to obtain the training set and the test set, the application adopts a random division method under a certain signal-to-noise ratio and a certain modulation method. That is, for each modulation type of signal, 80% of the data at each signal-to-noise ratio value is randomly selected as the training set, and 20% is randomly selected as the test set.

[0052] Step 4: Constructing a network model

[0053] The structure is built by using the Pytorch deep learning framework.

[0054] Step 5: Model training

[0055] The cross-entropy loss function is selected for the deep neural network loss function, and the Adam optimization function is selected for the deep neural network optimization function. The initial learning rate lr = 0.001 is selected. The equal interval adjustment learning rate StepLR is adopted for the learning rate modulation strategy, the adjustment interval is 50 Epochs, and the adjustment multiple is 0.1.

[0056] When training for 150 Epochs, the test set accuracy reaches the highest. The neural network parameter value file saved at 150 Epochs is selected and left for use when the deep neural network is in test mode.

[0057] Step 7: Model testing

[0058] The divided test set is input into the trained model, the output class of the model is compared with the label class of each data, and the accuracy of the model on the test set is calculated.

[0059] In order to evaluate the performance of the proposed one-dimensional complex residual network, the modulation classification accuracy of C-ResNet34, C-ResNet18, ResNet34, and ResNet18 under different signal-to-noise ratios is compared, as shown in Figure 1 At a relatively high signal-to-noise ratio (0-18db), the performance of the proposed complex model is improved by 3-10% compared with the corresponding real model. At a relatively low signal-to-noise ratio level (-20-0dB), the one-dimensional complex model of the present application produces a 2-2.2dB gain than the corresponding real network model at the same accuracy. This shows that the complex model can better learn the time-frequency domain features of the signal data and improve the noise resistance of the modulation recognition network.

[0060] At the same time, it is observed that C-ResNet34 performs better than C-ResNet18, which also shows that a deeper residual network can fit more complex function features, thereby having stronger feature extraction capability and achieving higher modulation recognition accuracy.

Claims

1. An electromagnetic signal identification method based on a one-dimensional complex-valued residual network, characterized in that... Includes the following steps: Step 1: Convert the communication signal received by the receiver into signal data in int16 format; Step 2: Perform a Fourier transform on the signal data from Step 1, plot the signal's spectrum, and estimate the carrier frequency of the signal. Step 3: Resample the signal data generated in Step 1 using four times the carrier frequency; Step 4: Perform intermediate frequency filtering on the signal obtained after resampling in Step 3 based on the carrier frequency estimated in Step 2; Step 5: Perform power normalization on the signal after intermediate frequency filtering in Step 4; Step 6: Divide the power-normalized dataset into a training set and a test set, design a one-dimensional complex-valued residual network, train the one-dimensional complex-valued residual network, and then use the test set to test the trained one-dimensional complex-valued residual network to obtain the corresponding digital signal modulation mode. The one-dimensional complex-valued residual network consists of 6 sub-modules, namely the first one-dimensional complex-valued convolution module, the first C-ResNet Block module, the second C-ResNet Block module, the third C-ResNet Block module, the fourth C-ResNet Block module, and the fully connected layer module. The 6 sub-modules are connected in sequence. Before training a one-dimensional complex-valued residual network, the weights must be initialized. Correct initialization of the one-dimensional complex-valued residual network is crucial for reducing the risk of vanishing or exploding gradients. The complex weights are represented as follows: W = Real{W} + iImag{W} When W is symmetrically distributed near 0, the variance of W is estimated from the parameter σ of the Rayleigh distribution. The weights are initialized according to the Rayleigh distribution, with an expected value of 0 and a variance parameter of 2σ. 2 The parameter σ will be set differently depending on the different neural network architectures; Step 7: Model Testing Input the pre-defined test set into the trained model, compare the model's output category with the label category of each data point, and calculate the model's accuracy on the test set.

2. The electromagnetic signal identification method based on a one-dimensional complex-valued residual network according to claim 1, characterized in that: The first C-ResNet Block module includes 3 complex-valued ResNet building blocks, the second C-ResNet Block module includes 4 complex-valued ResNet building blocks, the third C-ResNet Block module includes 6 complex-valued ResNet building blocks, and the fourth C-ResNet Block module includes 3 complex-valued ResNet building blocks.

3. The electromagnetic signal identification method based on a one-dimensional complex-valued residual network according to claim 2, characterized in that: The complex-valued ResNet building block includes a first one-dimensional complex-valued convolutional layer, a first one-dimensional complex Batchnorm layer, a first complex-valued ReLU activation function, a second one-dimensional complex-valued convolutional layer, a second one-dimensional complex-valued Batchnorm layer, a first complex-valued downsampled layer, and a second complex-valued ReLU activation function, with the above complex-valued convolutional layers and activation functions connected in series.

4. The electromagnetic signal identification method based on a one-dimensional complex-valued residual network according to claim 3, characterized in that: In the complex-valued convolutional layer, to simulate the complex-valued algorithm internally using the real-valued algorithm, a complex-valued convolutional kernel weight matrix W = A + iB and a complex-valued vector s = x + iy are defined, representing the I / Q signals respectively. The complex vector is then convolved using the obtained convolutional kernel. Where A and B are the real and imaginary parts of the complex convolution kernel weight matrix, respectively, and x and y are the real and imaginary parts of the complex vector, respectively. Then, the real and imaginary parts of the convolution result are obtained as follows: Each convolutional layer is followed by a complex batch normalization layer. Batch normalization is an important technique for optimizing data models. The real-valued complex batch normalization layer is represented as follows: in, V and V are the mean and variance of a batch of data, respectively; γ and β are trainable parameters, and ∈ is a very small value to prevent the denominator from being zero; However, for one-dimensional complex-valued residual networks: For complex batch normalization, learnable shift parameter β and scaling parameter γ are set, with scaling parameter γ given by the following formula: γ in the shift parameter β and the scaling parameter γ ri It will be initialized to zero, while the scaling parameter γ in γ will be initialized to zero. rr and γ ii Initialize to 5. The electromagnetic signal identification method based on a one-dimensional complex-valued residual network according to any one of claims 2-4, characterized in that: In the one-dimensional complex-valued residual network, the ResNet building block contains two mappings: a residual mapping and an identity mapping, wherein the residual mapping is described as follows: F(z)=H(z)-z This represents the expected underlying mapping of a ResNet block. In this structure, the network only needs to learn the differences between the input and output.

6. The electromagnetic signal identification method based on a one-dimensional complex-valued residual network according to claim 1, characterized in that: The one-dimensional complex-valued residual network adopts C-ResNet18 or C-ResNet34.

7. The electromagnetic signal identification method based on a one-dimensional complex-valued residual network according to claim 1, characterized in that: The loss function of the one-dimensional complex-valued residual network is the cross-entropy loss function, the optimization function is Adam, the initial learning rate is lr = 0.001, and the learning rate modulation strategy is StepLR with equal interval adjustment of the learning rate, with an adjustment interval of 50 epochs and an adjustment factor of 0.1.

Images