Modulation recognition method based on soft saturation IQ signal characteristics and complex-valued autoencoder

Abstract

The present invention discloses a modulation recognition method based on soft saturation IQ signal characteristics and complex-valued autoencoder, comprising the following steps: A: simulating a modulation signal, suppressing the impulse noise contained in the signal by using a soft saturation nonlinear function, and using it as a sample to prepare a training set and a test set; B: designing a complex-valued autoencoder and a signal classification network, and using the samples in the training set for pre-training, parameter migration, and signal classification; C: using the samples in the test set to test the modulation recognition method, and obtaining the modulation recognition accuracy by calculation. The present invention can make full use of the complementary information of the real part and the imaginary part of the signal in an impulse noise environment to improve the accuracy of modulation recognition.

Description

Technical Field

[0001] The invention belongs to the field of radio technology, and in particular relates to a modulation recognition method based on soft saturation IQ signal characteristics and a complex-valued autoencoder. Background Art

[0002] Modulation recognition technology can identify the modulation type of wireless communication signals in non-cooperative environments, which is crucial for the detection and demodulation of wireless signals. With the rapid advancement of wireless communication technology, modulation recognition is becoming increasingly important in many civil and military fields such as cognitive radio networks, network information security, and electronic warfare.

[0003] Modulation recognition methods can be mainly divided into two categories: methods based on maximum likelihood function and methods based on signal features. The rapid progress of deep learning theory and the significant increase in computer computing power have further promoted the research on modulation recognition methods based on signal features. Impulse noise is a very common noise. Its existence will seriously affect the accuracy of signal classification. At the same time, there is important complementary information between the real and imaginary vectors of the signal. Therefore, in order to achieve effective classification of signals in an impulsive noise environment, it is necessary to consider how to reduce the impact of impulse noise and utilize the joint action and intrinsic connection between the real and imaginary vectors of the signal. However, most of the existing modulation recognition methods using deep neural networks consider modulation recognition under Gaussian noise, and split the real and imaginary parts of the signal into two parts as the input of the neural network. Therefore, for modulation recognition methods, how to make full use of the complementary information of the real and imaginary parts of the signal in an impulsive noise environment to improve the accuracy of modulation recognition is an urgent problem to be solved. Summary of the invention

[0004] The purpose of the present invention is to provide a modulation recognition method based on soft saturation IQ signal characteristics and complex-valued autoencoder, which can fully utilize the complementary information of the real part and imaginary part of the signal in an impulse noise environment to improve the accuracy of modulation recognition.

[0005] To achieve the above object, the present invention provides a modulation recognition method based on soft saturation IQ signal characteristics and a complex-valued autoencoder, comprising the following steps:

[0006] A: Simulate the modulated signal, use the soft saturation nonlinear function to suppress the impulse noise contained in the signal, and use the suppression results as samples to create training sets and test sets;

[0007] B: Design a complex-valued autoencoder and a signal classification network, and use the samples in the training set for pre-training, parameter transfer, and signal classification;

[0008] C: Use the samples in the test set to test the modulation recognition method and calculate the modulation recognition accuracy.

[0009] As a further solution of the present invention: Step A specifically comprises the following steps:

[0010] A1: Get the modulated signal through simulation;

[0011] A2: Extract the real part I signal and the imaginary part Q signal from the modulated signal;

[0012] A3: Create a data set α, repeat steps A1 and A2 N times, prepare a modulation type label y for each signal while generating the signal, and use the simulated I-channel signal, Q-channel signal and the corresponding modulation type label y to form a data set α containing N signal samples;

[0013] A4: Prepare a data set β, use the soft saturation nonlinear function to perform data preprocessing on the real I-channel signal and the imaginary Q-channel signal of the signal in the data set α, and obtain a data set β containing N IQ signal samples processed by the soft saturation nonlinear function;

[0014] A5: Determine the training set and the test set, and divide the dataset α and the dataset β into the training set and the test set respectively according to the proportion.

[0015] As a further solution of the present invention: Step B specifically comprises the following steps:

[0016] B1: Design a complex-valued autoencoder, including an encoder and a decoder;

[0017] B2: Design a signal classification network, including a three-way feature extraction module;

[0018] B3: Perform pre-training and use complex-valued autoencoder for denoising;

[0019] B4: Perform parameter migration, replace the decoder of the complex-valued autoencoder in step B3 with the signal classification network designed in step B2, and use it as a classifier, retain the parameters of the pre-trained complex-valued autoencoder, and realize parameter migration;

[0020] B5: Perform signal classification.

[0021] As a further solution of the present invention: in step B1, the encoder is composed of 4 layers of modules, each layer includes a one-dimensional complex-valued convolution layer and a complex-valued activation function tanh; the decoder is composed of 4 layers of modules, each layer includes a one-dimensional complex-valued deconvolution layer and a complex-valued activation function tanh.

[0022] As a further solution of the present invention: in step B2, one path in the signal classification network is a spatial feature extraction module composed of a one-dimensional convolution and a one-dimensional zero-filling layer, the second path is a spatial feature extraction module composed of a two-dimensional convolution and a two-dimensional zero-filling layer, and the third path is a feature extraction module that cross-fuses the real and imaginary parts of the signal. The three paths of information are added together, and finally the signal classification is completed after the timing features are extracted through two layers of bidirectional gated recurrent units.

[0023] As a further solution of the present invention: the single training process in step B3 includes:

[0024] 1) Taking out the I-channel signal and the Q-channel signal contained in a sample in the training set of the data set α and forming a complex-valued input, wherein the I-channel signal and the Q-channel signal are respectively used as the real part and the imaginary part of the complex-valued input signal, and inputting the complex-valued input signal into the encoder of the complex-valued autoencoder in step B1, and finally obtaining the output of the encoder;

[0025] 2) Use the output of the encoder as the input of the decoder in step B1;

[0026] 3) The IQ signal features processed by the soft saturation nonlinear function in the data set β are used as pseudo-clean label data, and the parameters between the encoder input and the decoder output are optimized through the Huber loss function;

[0027] 4) Set a number of training rounds and establish an early stopping mechanism. When the loss of the test set of data set α does not decrease after multiple rounds, stop training until the loss function in step 3) converges and stabilizes, and the pre-training is completed.

[0028] As a further solution of the present invention: the process of signal classification in step B5 includes:

[0029] ① Extract the I-channel signal and the Q-channel signal contained in the IQ signal features processed by the soft saturation nonlinear function in the training set of the data set β and form a complex-valued input, and input the input signal to the encoder of the complex-valued autoencoder in step B1;

[0030] ②Use the output of the encoder as the input of the classifier;

[0031] ③ Use the cross entropy between the output of the classifier and the corresponding modulation type label y as the loss function for parameter optimization;

[0032] ④ Set a number of training rounds and establish an early stopping mechanism. If the loss of the training set does not decrease after multiple rounds, stop training until the loss function in step ③ converges and stabilizes.

[0033] As a further solution of the present invention: Step C specifically comprises the following steps:

[0034] C1: Use the signal samples in the test set of the data set β to test the proposed modulation recognition method. The I-channel signal and the Q-channel signal contained in all the signal samples in the test set of the data set β form a complex-valued input, and input the mixed signal classification network composed of the encoder part of the complex-valued autoencoder and the signal classifier in series. Load the parameters of the encoder part of the complex-valued autoencoder pre-trained in B3, and then use the output of the classifier as the modulation recognition result. And judge whether the modulation recognition result is correct according to the modulation type label y of the sample corresponding to the signal input to the encoder;

[0035] C2: The modulation recognition accuracy is obtained by calculation, and the average accuracy is calculated And use it as the modulation recognition accuracy, average accuracy Satisfies the following formula:

[0036]

[0037] Where N represents the total number of signal samples in the test set, M Acc Indicates the total number of correctly identified signal samples.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1) For impulse noise, a soft saturation nonlinear function is used to suppress the impulse noise in the signal. In the pre-training stage, a complex-valued autoencoder, a deep neural network, is designed for noise reduction and the structure of the complex-valued autoencoder is used for feature extraction. The parameters between the encoder input and the decoder output are optimized through the Huber loss function, and the influence of impulse noise is reduced through the designed complex-valued autoencoder.

[0040] 2) In the signal classification stage, the encoder part and the pre-trained parameters in the complex-valued autoencoder are retained to realize parameter migration; then the decoder part is replaced by the designed signal classification network; the loss function is formed by the cross entropy between the modulation type label y corresponding to the input signal feature and the output of the signal classification network, and the output information of the encoder after pre-training and denoising by the complex-valued autoencoder is extracted through the designed signal classification network to classify the signal, thereby improving the efficiency and accuracy of signal classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1It is an overall flow chart of the modulation identification method based on the soft saturation IQ signal feature and the complex-valued autoencoder of the present invention (I (In-phase): in-phase component, refers to the signal component with the same phase as the carrier signal, the I component carries the real part information of the signal, that is, the I-path signal; Q (Quadrature): quadrature component, refers to the signal component with a phase difference of 90 degrees (π / 2 radians) from the carrier signal, the Q component carries the imaginary part information of the signal, that is, the Q-path signal, which is independent of the I component. I+jQ represents a complex-valued signal formed by combining the I and Q components, where j is a complex-valued unit.)

[0042] Figure 2 This is the structural diagram of the complex-valued autoencoder in the present invention. (The autoencoder maps the input data to the latent space through its encoder part, and then maps the latent space features back to the original data space through the decoder part; in this process, the goal of the autoencoder is to minimize the reconstruction error, that is, the difference between the input data and the reconstructed data; in this way, the data representation learned by the autoencoder can capture the key features of the data and play a role in a variety of applications. Here, the latent space feature is the output of the encoder part of the complex-valued encoder)

[0043] Figure 3 It is the specific structure of the signal classification network in the present invention (V1, V2, K1, K2, Q1, Q2 are the outputs after two-dimensional convolution. Different names are given to distinguish them. In order to reflect the idea of ​​crossover and strengthen the connection between the real part and the imaginary part in the latent space features. R is a reshape transformation so that the features can be matrix multiplied or added. × represents the matrix multiplication between features)

[0044] Figure 4 The present invention provides experimental results and comparisons of the modulation recognition method based on soft saturation IQ signal characteristics and complex-valued autoencoder with some classical methods in the field of modulation recognition. DETAILED DESCRIPTION

[0045] The present invention will be further described below by way of examples.

[0046] like Figure 1 As shown, the modulation recognition method based on the soft saturation IQ signal feature and the complex-valued autoencoder comprises the following steps:

[0047] A: Simulate the modulated signal, use the soft saturation nonlinear function to suppress the impulse noise contained in the signal, and use the suppression results as samples to create training sets and test sets;

[0048] B: Design a complex-valued autoencoder and a signal classification network, and use the samples in the training set for pre-training, parameter transfer, and signal classification;

[0049] C: Use the samples in the test set to test the modulation recognition method and calculate the modulation recognition accuracy.

[0050] Further, step A specifically includes the following steps:

[0051] A1: The modulated signal is obtained through simulation; there are 11 modulation types involved, namely: BPSK, QPSK, 8PSK, 4PAM, GFSK, CPFSK, 16QAM, 64QAM, WBFM, AM-SSB, AM-DSB. The received modulated signal r(k) satisfies the following channel model formula:

[0052]

[0053] In the formula, a(k) represents the channel gain, x(k) represents the transmitted modulated signal, v(k) represents the pulse noise, and Ω represents the frequency offset. represents the phase shift, k represents the discrete time point, and the received signal can be represented by the vector r = [r 1 ,...,r K ] represents, K represents the signal length, e represents the exponential operation, and j is the complex value unit;

[0054] A2: Extract the real part of the signal I and the imaginary part of the signal Q from the modulated signal: I signal r (I) (k) and Q-path signal r (Q) (k) respectively satisfy the following formula:

[0055] r (I) (k) = Re[r(k)]

[0056] r (Q) (k) = Im[r(k)]

[0057] Where Re[·] represents the operation of taking the real part, and Im[·] represents the operation of taking the imaginary part.

[0058] A3: Create a data set α, repeat steps A1 and A2 N times, prepare a modulation type label y for each signal while generating the signal, and use the simulated I-channel signal, Q-channel signal and the corresponding modulation type label y to form a data set containing N signal samples. Wherein, the subscript n represents the number of simulation times, and a data set α containing IQ signals is obtained;

[0059] A4: Prepare a data set β, use the soft saturation nonlinear function to perform data preprocessing on the real I-channel signal and the imaginary Q-channel signal of the signal in the data set α, and obtain a data set β containing N IQ signal samples processed by the soft saturation nonlinear function; wherein the soft saturation nonlinear function satisfies the following formula:

[0060]

[0061] Where x represents the real part and imaginary part of the signal that need to be processed by the soft saturation nonlinear function respectively;

[0062] A5: Determine the training set and the test set, and divide the data set α and the data set β into the training set and the test set respectively in a ratio of 8:2.

[0063] Further, step B specifically includes the following steps:

[0064] B1: Design a complex-valued autoencoder, including an encoder and a decoder;

[0065] B2: Design a signal classification network, including a three-way feature extraction module;

[0066] B3: Perform pre-training and use complex-valued autoencoder for denoising;

[0067] B4: Perform parameter migration, replace the decoder of the complex-valued autoencoder in step B3 with the signal classification network designed in step B2, and use it as a classifier, retain the parameters of the pre-trained complex-valued autoencoder, and realize parameter migration;

[0068] B5: Perform signal classification.

[0069] Further, such as Figure 2 As shown, in step B1, the encoder consists of 4 layers of modules, each of which includes a one-dimensional complex-valued convolution layer and a complex-valued activation function tanh; the decoder consists of 4 layers of modules, each of which includes a one-dimensional complex-valued deconvolution layer and a complex-valued activation function tanh.

[0070] Further, such as Figure 3 As shown, in step B2, one path in the signal classification network is a spatial feature extraction module composed of a one-dimensional convolution and a one-dimensional zero-filling layer, the second path is a spatial feature extraction module composed of a two-dimensional convolution and a two-dimensional zero-filling layer, and the third path is a feature extraction module that cross-fuses the real and imaginary parts of the signal. The three paths of information are added together, and finally the signal classification is completed after the timing features are extracted through two layers of bidirectional gated recurrent units.

[0071] Furthermore, the single training process in step B3 includes:

[0072] 1) Taking out the I-channel signal and the Q-channel signal contained in a sample in the training set of the data set α and forming a complex-valued input, wherein the I-channel signal and the Q-channel signal are respectively used as the real part and the imaginary part of the complex-valued input signal, and inputting the complex-valued input signal into the encoder of the complex-valued autoencoder in step B1, and finally obtaining the output of the encoder;

[0073] 2) Use the output of the encoder as the input of the decoder in step B1;

[0074] 3) The IQ signal features processed by the soft saturation nonlinear function in the data set β are used as pseudo-clean label data, and the parameters between the encoder input and the decoder output are optimized through the Huber loss function;

[0075] 4) Set a number of training rounds, preferably 300 times, and set up an early stopping mechanism. When the loss of the test set of data set α does not decrease for multiple rounds, the training is stopped, preferably 15 rounds, until the loss function in step 3) converges and stabilizes, and the pre-training is completed. During the pre-training process, Adam is used as the optimizer of the network parameters, and the learning rate is set to 0.0001.

[0076] Furthermore, the signal classification process in step B5 includes:

[0077] ① Extract the I-channel signal and the Q-channel signal contained in the IQ signal features processed by the soft saturation nonlinear function in the training set of the data set β and form a complex-valued input, and input the input signal to the encoder of the complex-valued autoencoder in step B1;

[0078] ②Use the output of the encoder as the input of the classifier;

[0079] ③ Use the cross entropy between the output of the classifier and the corresponding modulation type label y as the loss function for parameter optimization;

[0080] ④ Set a number of training rounds, preferably 300 times, and set up an early stopping mechanism. Stop training when the loss of the training set does not decrease after multiple rounds, preferably 15 rounds, until the loss function in step ③ converges and stabilizes. Use Adam as the optimizer for network parameters, and set the learning rate to 0.001.

[0081] Further, step C specifically includes the following steps:

[0082] C1: Use the signal samples in the test set of the data set β to test the proposed modulation recognition method. The I-channel signal and the Q-channel signal contained in all the signal samples in the test set of the data set β form a complex-valued input, and input the mixed signal classification network composed of the encoder part of the complex-valued autoencoder and the signal classifier in series. Load the parameters of the encoder part of the complex-valued autoencoder pre-trained in B3, and then use the output of the classifier as the modulation recognition result. And judge whether the modulation recognition result is correct according to the modulation type label y of the sample corresponding to the signal input to the encoder;

[0083] C2: The modulation recognition accuracy is obtained by calculation, and the average accuracy is calculated And use it as the modulation recognition accuracy, average accuracy Satisfies the following formula (Acc is the abbreviation of Accuracy):

[0084]

[0085] Where N represents the total number of signal samples in the test set, M Acc Indicates the total number of correctly identified signal samples.

[0086] like Figure 4 As shown, the "complex-valued autoencoder" curve is the modulation recognition accuracy obtained by the modulation recognition method based on soft saturation IQ signal characteristics and complex-valued autoencoder proposed in the present invention, and the other comparison curves are some classic methods in the field of modulation recognition. The results prove the superiority of the modulation recognition method based on soft saturation IQ signal characteristics and complex-valued autoencoder in an impulse noise environment proposed in this paper.

Claims

1. A modulation recognition method based on soft saturation IQ signal characteristics and complex-valued autoencoder, characterized in that: The following steps are involved: A: Simulate the modulated signal, use the soft saturation nonlinear function to suppress the impulse noise contained in the signal, and use the suppression results as samples to create training sets and test sets; Step A specifically includes the following steps: A1: Get the modulated signal through simulation; A2: Extract the real part I signal and the imaginary part Q signal from the modulated signal; A3: Create a data set α, repeat steps A1 and A2 N times, prepare a modulation type label y for each signal while generating the signal, and use the simulated I-channel signal, Q-channel signal and the corresponding modulation type label y to form a data set α containing N signal samples; A4: Prepare a data set β, use the soft saturation nonlinear function to perform data preprocessing on the real I-channel signal and the imaginary Q-channel signal of the signal in the data set α, and obtain a data set β containing N IQ signal samples processed by the soft saturation nonlinear function; A5: Determine the training set and the test set, and divide the dataset α and the dataset β into the training set and the test set respectively in a ratio of 8:2; B: Design a complex-valued autoencoder and a signal classification network, and use the samples in the training set for pre-training, parameter transfer, and signal classification; Step B specifically includes the following steps: B1: Design a complex-valued autoencoder, including an encoder and a decoder; B2: Design a signal classification network, including a three-way feature extraction module; In step B2, one path in the signal classification network is a spatial feature extraction module composed of a one-dimensional convolution and a one-dimensional zero padding layer, the second path is a spatial feature extraction module composed of a two-dimensional convolution and a two-dimensional zero padding layer, and the third path is a feature extraction module that cross-fuses the real part and the imaginary part of the signal. The three paths of information are added, and finally the signal classification is completed after the timing features are extracted through a two-layer bidirectional gated recurrent unit; B3: Perform pre-training and use complex-valued autoencoder for denoising; B4: Perform parameter migration, replace the decoder of the complex-valued autoencoder in step B3 with the signal classification network designed in step B2, and use it as a classifier, retain the parameters of the pre-trained complex-valued autoencoder, and realize parameter migration; B5: perform signal classification; C: Use the samples in the test set to test the modulation recognition method and calculate the modulation recognition accuracy; Step C specifically includes the following steps: C1: Use the signal samples in the test set of the data set β to test the proposed modulation recognition method. The I-channel signal and the Q-channel signal contained in all the signal samples in the test set of the data set β form a complex-valued input, and input the mixed signal classification network composed of the encoder part of the complex-valued autoencoder and the signal classifier in series. Load the parameters of the encoder part of the complex-valued autoencoder pre-trained in B3, and then use the output of the classifier as the modulation recognition result. And judge whether the modulation recognition result is correct according to the modulation type label y of the sample corresponding to the signal input to the encoder; C2: The modulation recognition accuracy is obtained by calculation, and the average accuracy is calculated And use it as the modulation recognition accuracy, average accuracy Satisfies the following formula: Where N represents the total number of signal samples in the test set, M Acc Indicates the total number of correctly identified signal samples.

2. The modulation recognition method based on soft saturation IQ signal characteristics and complex-valued autoencoder according to claim 1, characterized in that: In step B1, the encoder consists of 4 layers of modules, each of which contains a one-dimensional complex-valued convolutional layer and a complex-valued activation function tanh; The decoder consists of 4 layers of modules, each of which contains a one-dimensional complex-valued deconvolution layer and a complex-valued activation function tanh.

3. The modulation recognition method based on soft saturation IQ signal characteristics and complex-valued autoencoder according to claim 1, characterized in that: The single training process in step B3 includes: 1) Taking out the I-channel signal and the Q-channel signal contained in a sample in the training set of the data set α and forming a complex-valued input, wherein the I-channel signal and the Q-channel signal are respectively used as the real part and the imaginary part of the complex-valued input signal, and inputting the complex-valued input signal into the encoder of the complex-valued autoencoder in step B1, and finally obtaining the output of the encoder; 2) Use the output of the encoder as the input of the decoder in step B1; 3) The IQ signal features processed by the soft saturation nonlinear function in the data set β are used as pseudo-clean label data, and the parameters between the encoder input and the decoder output are optimized through the Huber loss function; 4) Set a number of training rounds and establish an early stopping mechanism. When the loss of the test set of data set α does not decrease after multiple rounds, stop training until the loss function in step 3) converges and stabilizes, and the pre-training is completed.

4. The modulation recognition method based on soft saturation IQ signal characteristics and complex-valued autoencoder according to claim 1, characterized in that: The signal classification process in step B5 includes: ① Extract the I-channel signal and the Q-channel signal contained in the IQ signal features processed by the soft saturation nonlinear function in the training set of the data set β and form a complex-valued input, and input the input signal to the encoder of the complex-valued autoencoder in step B1; ②Use the output of the encoder as the input of the classifier; ③ Use the cross entropy between the output of the classifier and the corresponding modulation type label y as the loss function for parameter optimization; ④ Set a number of training rounds and establish an early stopping mechanism. If the loss of the training set does not decrease after multiple rounds, stop training until the loss function in step ③ converges and stabilizes.

Images