Communication signal modulation recognition method based on residual neural network and meta-learning fusion

Abstract

The present invention proposes a communication signal modulation recognition method based on the fusion of residual neural network and meta-learning, which belongs to the field of communication technology and includes the following methods: communication modulation signal acquisition; subtask division; residual neural network construction; sample extraction; updating residual neural network parameters; updating meta-learning loss function; updating residual neural network initial parameters; iteration; determining residual neural network parameters for test signals; inputting the final test communication signal into the residual neural network to obtain the final classification result. The present invention integrates the residual neural network with the meta-learning method, which can effectively learn the good features of the signal, solve the problem of weak generalization ability of the network model, and only requires a small number of samples to complete the signal recognition.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular relates to a communication signal modulation recognition method based on the fusion of residual neural network and meta-learning. Background Art

[0002] Wireless signal recognition (WSR) has broad prospects in military and civilian applications, which may include signal reconnaissance and interception, anti-interference, and device identification. Generally, WSR mainly includes modulation recognition (MR) and wireless technology recognition (WTR). MR, also known as automatic modulation classification (AMC), was first widely used in the military field and then expanded to the civilian field. MR classifies radio signals by identifying modulation patterns, which helps to evaluate wireless transmission schemes and device types. Moreover, MR is able to extract digital baseband information even when prior information is limited.

[0003] Traditional modulation recognition algorithms can be mainly divided into two categories: likelihood-based (LB) and feature-based (FB) methods. The LB method is based on hypothesis testing theory; the decision-theory-based method has the best performance but high computational complexity. Therefore, feature-based methods were developed as suboptimal classifiers for practical applications. However, feature-based methods require manual feature extraction and are more complex to implement. In recent years, due to the outstanding performance of deep learning in computer vision, machine translation, speech recognition, and natural language processing, the application of deep learning and machine learning in the field of communications has gradually attracted attention.

[0004] Machine learning algorithms in the field of communication have the characteristics of not requiring special feature extraction, flexible models, and end-to-end learning. However, when the communication signal is greatly affected by noise, the recognition rate of the communication signal will be seriously affected. When the network depth of the deep learning model is increased to improve the recognition effect, "degeneration" often occurs, that is, as the network depth continues to increase, the model accuracy decreases significantly without any warning. How to ensure the accuracy of the model while deepening the network depth is a problem that needs to be solved. Secondly, the machine learning model is a data-driven model that requires a large amount of data to ensure the implementation effect of the model. When the available data is scarce, modern deep neural networks often fail to achieve the desired effect. Therefore, effectively learning the characteristics of data through a small number of samples is also a problem to be solved. Thirdly, the current machine learning model mostly uses random initialization to initialize parameters. How to use some prior knowledge to initialize the parameters so that the model understands the characteristics of the data and identifies and distinguishes specific data in a targeted manner is also a problem worth studying. Finally, the current machine learning algorithm is greatly affected by data quality in the case of pure data drive. It is difficult to remain robust in the presence of imperfect data (such as missing values ​​or noise values, outliers, etc.), which leads to weak model generalization ability. Machine learning algorithms need to improve their generalization capabilities to lead to effective implementation of the algorithms in practical applications. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art, such as low recognition rate when the signal is affected by noise, easy degradation of the network model, weak generalization ability of the network model, and great influence of the amount of data, and to provide a communication signal modulation recognition method based on the fusion of residual neural network and meta-learning, which uses the special structure of the residual neural network to effectively solve the "degradation" problem and effectively improve the signal recognition rate; in addition, the meta-learning method is used to enable the model to learn adaptive initialization parameters using different noises. By effectively initializing the network parameters, the model can quickly adapt to the modulation recognition of communication signals. This method can effectively learn the good characteristics of the signal, solve the problem of weak generalization ability of the network model, and only a small number of samples are needed to complete the signal recognition.

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: a communication signal modulation recognition method based on the fusion of residual neural network and meta-learning, comprising the following steps:

[0007] Step 1: Acquire the communication modulation signal

[0008] A variety of communication modulation signals with different signal-to-noise ratios under additive white Gaussian noise are obtained as training data sets. The expressions of communication modulation signals with different signal-to-noise ratios under additive white Gaussian noise are:

[0009]

[0010] Where s[t] and r[t] represent the transmission signal and the received signal in the tth time period, respectively; γ[t] is the communication channel gain; ω0 is the carrier frequency offset; θ0 is the phase offset between the sender and the receiver; v[t] is the additive white Gaussian noise; and j is the imaginary unit.

[0011] The expression of signal-to-noise ratio is:

[0012]

[0013] Among them, P r is the power of the signal, P v is the power of the noise.

[0014] Step 2: Subtask Division

[0015] The training data set is divided into multiple subtask data according to the number of noise signal-to-noise ratios. The signal recognition under one signal-to-noise ratio is one subtask. The number of noise signal-to-noise ratios is N, and the number of subtasks is N. The subtasks are numbered according to the size of the signal-to-noise ratio. The larger the signal-to-noise ratio, the larger the number. Each subtask data contains r types of communication signals, and the types of communication signals contained in each subtask data are the same. The number of communication signal types contained in each subtask data is the same as the number of communication modulation signal types used in the final test, and the number of samples in each subtask is the same.

[0016] Step 3: Residual Neural Network Construction

[0017] A residual neural network structure is constructed for signal classification. The residual neural network consists of 6 residual blocks, 2 fully connected layers, 1 Dropout layer and 1 Softmax layer. The fully connected layer integrates the automatically extracted high-level features into local feature information.

[0018] The Softmax layer is used for multi-class classification and its expression is:

[0019]

[0020] Among them, a k' is the k'th input signal of the output layer, y k' is the output of the k'th neuron, C is the number of output nodes, that is, the number of classification categories, a c Represents the cth input signal of the output layer.

[0021] Each residual block adopts the network structure of residual block 1 or residual block 2. The residual block 1 and residual block 2 are composed of a convolution layer, two sub-residual blocks and a pooling layer. Each sub-residual block contains 2 convolution layers, and all convolution layers contain 32 convolution kernels (including the convolution layer in the sub-residual block and the convolution layer outside the sub-residual block), with a step size of 1 and a padding of 1. The convolution kernel sizes of the sub-residual blocks of residual block 1 and residual block 2 are different. The convolution kernel size of the sub-residual block of residual block 1 is 3×2, and the convolution kernel size of the sub-residual block of residual block 2 is 3×1. The convolution layer outside the sub-residual block is activated by the Relu function. The output of each sub-residual block is also activated by the Relu function and maximum pooling is performed. The convolution layer performs a convolution operation on the input to extract the characteristics of the signal. The expression of the convolution operation is:

[0022] x l′ =f(w l′ *x l′-1 +b l′ )

[0023] Among them, l' represents the number of network layers (network layers include input layers and convolutional layers), w l′ represents the convolution kernel parameters in the current convolution layer, b l′ is the bias vector in the current convolutional layer, f(·) represents the activation function, x l′-1 Represents the output of the previous network layer (the network layer includes the input layer and the convolutional layer), x l′ Represents the output of the current convolutional layer;

[0024] The pooling layer downsamples the convolutional layer. The expression of the pooling layer is:

[0025] x l′ =down(x l′-1 )

[0026] Where down(·) represents the pooling function; x l′-1 represents the output of the previous convolutional layer, x l′ Represents the output of the current pooling layer;

[0027] The Relu function expression is:

[0028]

[0029] Among them, f(x) is the function that Relu takes the maximum value. When the input is negative, the output is 0, the neuron will not be activated, and x is the input of the function.

[0030] Step 4: Extract samples

[0031] Randomly extract samples from the subtask data and divide them into training sets and test sets as the input of the residual neural network. The number of samples randomly extracted from the subtask data each time is h+m, and the samples contain r types of communication signals, of which h samples are used as the training set of the extracted samples, and m samples are used as the test set of the extracted samples.

[0032] The input data of the residual neural network is n in ×l×dim×1,n in is the number of input samples of the residual neural network, l is the signal sampling length. The signal is in complex form, and the real part and the imaginary part of the signal are superimposed together, so dim=2.

[0033] Step 5: Update the residual neural network parameters

[0034] The initial parameters of the residual neural network are randomly initialized, and the extracted sample training set is input into the residual neural network, and the residual neural network parameters are updated through rapid iteration.

[0035] The steps to update the residual neural network parameters are:

[0036] 1) Randomly initialize the residual neural network parameters φ0;

[0037] 2) Input the extracted sample training set into the residual neural network;

[0038] 3) Define the cross entropy loss function l(φ) using the extracted sample labels and network output results;

[0039] Find the gradient of this function and update the parameters. Among them, k is the number of samples drawn, α is the inner loop learning rate, φ k is the residual neural network parameter when the kth sample is extracted, l(φ k ) is the cross entropy loss function of the network parameters, are the residual neural network parameters after gradient update.

[0040] The cross entropy loss function is defined as:

[0041] l(φ)=[y (k) logg(x (k) )+(1-y (k) )log(1-g(x (k) ))]

[0042] Among them, x (k) is the data of the sample training set extracted for the kth time, y (k) is the label of the data of the k-th sample training set, g represents the residual neural network structure, and l(φ) is the cross entropy loss function.

[0043] Iteratively updating parameters must ensure that the residual neural network performs well on the extracted data training set. The condition for judging whether it performs well is: when the cross entropy loss function l(φ) decreases for a period of time, it continues to iterate, and l(φ) remains stable, which proves that the residual neural network performs well on the extracted data training set.

[0044] Step 6: Update the meta-learning loss function

[0045] The sample test set is input into the updated network to obtain the output data corresponding to the test set. The expression for updating the meta-learning loss function based on the labels corresponding to the test set and the output data corresponding to the test set is:

[0046]

[0047] in, is the data of the sample test set extracted for the kth time, is the label of the data of the sample test set extracted for the kth time, is the meta-learning loss function updated after k-1 samples are drawn,

[0048] Step 7: Update the initial parameters of the residual neural network

[0049] The gradient of the meta-learning loss function is calculated, and the initial parameters of the residual neural network are updated using this gradient. The expression for updating the initial parameters of the residual neural network using the gradient is:

[0050]

[0051] Among them, β is the outer loop learning rate. The inner loop learning rate α should be greater than the outer loop learning rate β. is the updated initial parameter of the residual neural network.

[0052] Step 8. Iteration

[0053] Repeat steps 4 to 7 above multiple times. When repeating step 5, the parameters of the residual neural network are no longer randomly initialized, but are initialized to the network parameters updated in the previous iteration.

[0054] Step 9: Determine the residual neural network parameters for the test signal

[0055] Input the communication modulation signal with the same properties as the final test signal into the residual neural network, repeat step 5, and adjust the initial parameters of the residual neural network. When repeating step 5, the parameters of the residual neural network are no longer randomly initialized, but are initialized to the network parameters after the last iteration update.

[0056] Step 10: Output the results

[0057] The final test communication signal is input into the residual neural network to obtain the final classification result.

[0058] Through the above design scheme, the present invention can bring the following beneficial effects:

[0059] 1. The present invention proposes a communication signal modulation recognition method based on a residual neural network. By utilizing the special structure of the residual block (shortcut connection method), the difference between the target value and the residual block input, i.e., the residual, is learned so that the accuracy no longer decreases as the network deepens. This can solve the "degeneration" problem when the network depth increases, and improve the recognition accuracy of communication signal modulation.

[0060] 2. The present invention also proposes a communication signal modulation recognition method based on meta-learning. By utilizing the prior knowledge of the communication signal, the initial parameters of the neural network are initialized. After learning, the initialization parameters of the neural network contain high-quality features of the signal, which can ensure that the effective features of the signal can be learned without a large amount of data, but only a few samples.

[0061] 3. The communication signal modulation recognition method based on meta-learning described in the present invention utilizes the prior knowledge of signal-to-noise ratio and communication signal, and learns to initialize the communication signal classification model by extracting signals under different signal-to-noise ratios, so that the model can adapt to signals under different signal-to-noise ratios and can adapt to the recognition of different communication signals under different noises, thereby improving the generalization ability of the classification model. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a flowchart of a communication signal modulation recognition method based on the fusion of residual neural network and meta-learning of the present invention;

[0063] Figure 2 It is a signal time domain waveform diagram of a -5dB 4ASK signal in an embodiment of a communication signal modulation identification method based on the fusion of residual neural network and meta-learning of the present invention;

[0064] Figure 3 It is a signal time domain waveform diagram of a -5dB 4FSK signal in an embodiment of the communication signal modulation recognition method based on the fusion of residual neural network and meta-learning of the present invention;

[0065] Figure 4 It is a signal time domain waveform diagram of a -5dB 4PSK signal in an embodiment of the communication signal modulation identification method based on the fusion of residual neural network and meta-learning of the present invention;

[0066] Figure 5 It is a signal time domain waveform diagram of a 4QAM signal -5dB in an embodiment of the communication signal modulation identification method based on the fusion of residual neural network and meta-learning of the present invention;

[0067] Figure 6 It is a signal time domain waveform diagram of 4ASK signal 0dB in an embodiment of the communication signal modulation recognition method based on the fusion of residual neural network and meta-learning of the present invention;

[0068] Figure 7 It is a signal time domain waveform diagram of 4FSK signal 0dB in an embodiment of the communication signal modulation recognition method based on residual neural network and meta-learning fusion of the present invention;

[0069] Figure 8 It is a signal time domain waveform diagram of 4PSK signal 0dB in an embodiment of the communication signal modulation identification method based on the fusion of residual neural network and meta-learning of the present invention;

[0070] Fig. 9 It is a signal time domain waveform diagram of 4QAM signal 0dB in an embodiment of the communication signal modulation identification method based on the fusion of residual neural network and meta-learning of the present invention;

[0071] Fig.10 It is a signal time domain waveform diagram of 4ASK signal 5dB in an embodiment of the communication signal modulation recognition method based on the fusion of residual neural network and meta-learning of the present invention;

[0072] Fig.11 It is a signal time domain waveform diagram of 4FSK signal 5dB in an embodiment of the communication signal modulation recognition method based on the fusion of residual neural network and meta-learning of the present invention;

[0073] Fig.12 It is a signal time domain waveform diagram of 4PSK signal 5dB in an embodiment of the communication signal modulation recognition method based on residual neural network and meta-learning fusion of the present invention;

[0074] Fig.13 It is a signal time domain waveform diagram of 4QAM signal 5dB in an embodiment of the communication signal modulation identification method based on the fusion of residual neural network and meta-learning of the present invention;

[0075] Fig.14 It is a flow chart of decomposing training subtasks in an embodiment of a communication signal modulation recognition method based on the fusion of residual neural network and meta-learning of the present invention;

[0076] Fig.15 It is a residual neural network structure diagram in an embodiment of the communication signal modulation recognition method based on the fusion of residual neural network and meta-learning of the present invention;

[0077] Fig.16 It is a structural diagram of a residual block in an embodiment of a communication signal modulation recognition method based on the fusion of residual neural network and meta-learning of the present invention;

[0078] Fig.17 It is a structural diagram of residual block 2 in an embodiment of the communication signal modulation recognition method based on the fusion of residual neural network and meta-learning of the present invention;

[0079] Fig.18 A flow chart of updating residual neural network parameters in an embodiment of a communication signal modulation recognition method based on the fusion of residual neural network and meta-learning of the present invention;

[0080] Fig.19 A flow chart of updating initial parameters of a residual neural network in an embodiment of a communication signal modulation recognition method based on the fusion of a residual neural network and meta-learning according to the present invention;

[0081] Fig. 20 This is a test task flow chart in an embodiment of the communication signal modulation recognition method based on the fusion of residual neural network and meta-learning in the present invention. DETAILED DESCRIPTION

[0082] In order to explain the present invention more clearly, the present invention is further described below in conjunction with preferred embodiments and accompanying drawings. It should be understood by those skilled in the art that the content described below is illustrative rather than restrictive, and should not be used to limit the scope of protection of the present invention. Unless otherwise defined, the technical terms or scientific terms used herein should be the common meanings understood by people with ordinary skills in the field to which the present invention belongs.

[0083] like Figure 1 As shown, the present invention relates to a communication signal modulation recognition method based on the fusion of residual neural network and meta-learning. First, communication modulation signals with different signal-to-noise ratios under additive Gaussian white noise are obtained as training objects, and the entire data set is divided into multiple sub-task data sets according to different signal-to-noise ratios of the signal. The recognition of the signal under one signal-to-noise ratio corresponds to one sub-task, and the sub-tasks are numbered accordingly; secondly, a residual neural network is constructed for signal classification; then, using the meta-learning theory, the initial parameters of the residual neural network are obtained through learning: randomly sampled data in each sub-task data is used as the input of the residual neural network, and each sub-task is sampled until all sub-tasks are used, so that the network uses the characteristics of the signal to learn how to initialize the parameters; thirdly, a part of the communication modulation signals with known labels and the same properties as the final test signals are obtained, and they are input into the residual neural network for training to complete the parameter update of the residual neural network; finally, the communication modulation signal used for the final test is obtained, and it is input into the residual neural network to complete the communication signal modulation recognition task. The present invention integrates the residual neural network with the meta-learning method, avoids the "degeneration" phenomenon that occurs during the training process, improves the generalization ability of the neural network, and improves the accuracy of modulation recognition.

[0084] The specific implementation process is as follows:

[0085] 1. Use Matlab software to simulate and generate training and test task data sets. The data sets contain 4ASK, 4FSK, 4PSK, and 4QAM signals. The signal-to-noise ratio range is -5dB to 5dB, and the noise is additive white Gaussian noise. The expression of the communication modulation signal under additive white Gaussian noise with different signal-to-noise ratios is:

[0086]

[0087] Where s[t] and r[t] represent the transmission signal and the received signal in the tth time period, respectively; γ[t] is the communication channel gain; ω0 is the carrier frequency offset; θ0 is the phase offset between the sender and the receiver; v[t] is the additive white Gaussian noise; and j is the imaginary unit.

[0088] The expression of signal-to-noise ratio is:

[0089]

[0090] Among them, P r is the power of the signal, P v is the power of the noise. Figure 2 to Figure 5 This is the time domain waveform of the -5dB signal generated by the simulation. Figure 6 to Figure 9 This is the time domain waveform of the 0dB signal generated by the simulation. Figure 10 to Figure 13 This is the time domain waveform of the 5dB signal generated by the simulation.

[0091] The training task is divided into 11 subtasks according to the number of noise signal-to-noise ratios, and each subtask contains four types of communication signals generated. The decomposition diagram of the training subtasks is shown in Fig.14 shown.

[0092] Second, under the premise that the neural network can converge, as the network depth increases, the network performance first gradually increases to saturation, and then rapidly decreases. This is the "degradation" phenomenon, which makes it difficult to improve the recognition rate of classification problems. The residual neural network uses its special structure to enable the input signal to be directly transmitted from any low layer to the high layer during forward propagation of the neural network. Because it contains a natural identity mapping, it can solve the network degradation problem to a certain extent.

[0093] A residual neural network structure is constructed for signal classification. The residual neural network consists of 6 residual blocks, 2 fully connected layers, 1 Dropout layer and 1 Softmax layer. The fully connected layer integrates the automatically extracted high-level features into local feature information.

[0094] The Softmax layer is used for multi-class classification and its expression is:

[0095]

[0096] Among them, a k' is the k'th input signal of the output layer, y k' is the output of the k'th neuron, C is the number of output nodes, that is, the number of classification categories, a c Represents the cth input signal of the output layer.

[0097] Each residual block adopts the network structure of residual block one or residual block two, such as Fig.15 As shown in the figure, the six residual blocks respectively use residual module 1, residual module 2, residual module 2, residual module 2, residual module 2 and residual module 2 in sequence; the residual block 1 and residual block 2 are composed of a convolution layer, two sub-residual blocks and a pooling layer, each sub-residual block contains 2 convolution layers, all convolution layers contain 32 convolution kernels, the step size is 1, and the padding is 1; the sub-residual block convolution kernel sizes of residual block 1 and residual block 2 are different, the sub-residual block convolution kernel size of residual block 1 is 3×2, and the sub-residual block convolution kernel size of residual block 2 is 3×1; the convolution layer outside the sub-residual block is activated by the Relu function; the output of each sub-residual block is also activated by the Relu function and maximum pooling is performed, wherein the convolution layer performs a convolution operation on the input to extract the characteristics of the signal, and the expression of the convolution operation is:

[0098] x l′ =f(w l′ *x l′-1 +b l′ )

[0099] Among them, l' represents the number of network layers (network layers include input layers and convolutional layers), w l′ represents the convolution kernel parameters in the current convolution layer, b l′ is the bias vector in the current convolutional layer, f(·) represents the activation function, x l′-1 Represents the output of the previous network layer (the network layer includes the input layer and the convolutional layer), x l′ Represents the output of the current convolutional layer.

[0100] The pooling layer downsamples the convolutional layer. The expression of the pooling layer is:

[0101] x l′ =down(x l′-1 )

[0102] Where down(·) represents the pooling function. l′-1 represents the output of the previous convolutional layer, x l′ Represents the output of the current pooling layer.

[0103] The Relu function expression is:

[0104]

[0105] Among them, f(x) is the function that Relu takes the maximum value. When the input is negative, the output is 0, the neuron will not be activated, and x is the input of the function.

[0106] The completed residual neural network structure is as follows: Fig.15 The two residual block structures in the residual neural network are shown in Fig.16 and Fig.17 shown.

[0107] 3. Randomly extract samples from the subtask and divide them into training sets and test sets as the input of the residual neural network.

[0108] The initial parameters of the residual neural network are randomly initialized, and the extracted sample training set is input into the residual neural network, and the residual neural network parameters are updated through rapid iteration.

[0109] The steps to update the residual neural network parameters are:

[0110] 1) Randomly initialize the residual neural network parameters φ0;

[0111] 2) Input the extracted sample training set into the residual neural network;

[0112] 3) Define the cross entropy loss function l(φ) using the extracted sample labels and network output results;

[0113] Find the gradient of this function and update the parameters. Among them, k is the number of samples drawn, α is the inner loop learning rate, φ k is the residual neural network parameter when the kth sample is extracted, l(φ k ) is the cross entropy loss function of the network parameters, are the residual neural network parameters after gradient update.

[0114] The cross entropy loss function is defined as:

[0115] l(φ)=[y (k) logg(x (k) )+(1-y (k) )log(1-g(x (k) ))]

[0116] Among them, x (k) is the data of the sample training set extracted for the kth time, y (k) is the label of the data of the k-th sample training set, and g represents the residual neural network structure.

[0117] Iteratively updating parameters must ensure that the residual neural network performs well on the extracted data training set. The condition for judging whether it performs well is: when the cross entropy loss function After falling for a period of time, continue to iterate. It remains stable, proving that the residual neural network performs well on the extracted data training set.

[0118] The flowchart of the residual neural network updating parameters is as follows Fig.18 shown.

[0119] Fourth, input the sample test set into the updated residual neural network to obtain the output data corresponding to the test set. The expression for updating the meta-learning loss function based on the labels corresponding to the test set and the output data corresponding to the test set is:

[0120]

[0121] in, is the data of the sample test set extracted for the kth time, is the label of the data of the sample test set extracted for the kth time, is the meta-learning loss function updated after k-1 samples are drawn,

[0122] The gradient of the meta-learning loss function is calculated, and the initial parameters of the residual neural network are updated using this gradient. The expression for updating the initial parameters of the residual neural network using the gradient is:

[0123]

[0124] Among them, β is the outer loop learning rate, and the inner loop learning rate α must be greater than the outer loop learning rate β; is the updated initial parameter of the residual neural network.

[0125] Randomly draw samples multiple times and repeat the above steps. The flowchart for updating the initial parameters of the residual neural network is as follows: Fig.19As shown. Extract sample data of subtasks and divide the data into subtask training set and subtask test set. Input the subtask training set into the residual neural network and update the residual neural network parameters. Then input the subtask test set into the residual neural network after updating the parameters, and use the subtask test set label and the output of the subtask test set input into the residual neural network to update the meta-learning loss function. By calculating the gradient of the meta-learning loss function, update the initial parameters of the residual neural network. By extracting subtask data multiple times, the parameters of the residual neural network are initialized. Updating the initial parameters of the residual neural network can make the residual neural network adapt to signals under different signal-to-noise ratios, effectively learn the key features of the signal, and suppress the interference of noise to a certain extent. When a new task appears, the residual neural network can adjust its parameters according to the characteristics of the new task so that it performs well on this task.

[0126] Test flow chart as follows Fig. 20 As shown. First, a signal with the same properties as the test signal is input into the residual neural network, and the residual neural network parameters are updated so that the residual neural network is suitable for the test data. Finally, the test communication signal is input into the residual neural network with updated parameters to obtain the final classification result.

Claims

1. A communication signal modulation recognition method based on residual neural network and meta-learning fusion, characterized in that: The steps include: Step 1: Obtain a variety of communication modulation signals with different signal-to-noise ratios under additive white Gaussian noise as training data sets; Step 2: Divide the training data set into N subtask data according to the number of noise signal-to-noise ratios. The recognition of a signal under a signal-to-noise ratio corresponds to a subtask, and the subtasks are numbered according to the size of the signal-to-noise ratio. The larger the signal-to-noise ratio, the larger the number. Each subtask contains multiple communication signals. Step 3: Construct a residual neural network structure for signal classification; Step 4: Randomly select a small number of samples less than 10% of the subtask data from the lowest-numbered subtask data in the subtask that has not been used currently and divide them into training sets and test sets as the input of the residual neural network; Step 5: Randomly initialize the initial parameters of the residual neural network, and input the extracted sample training set into the residual neural network, quickly iterate, and update the residual neural network parameters; Step 6: Input the extracted sample test set into the updated residual neural network, and update the meta-learning loss function according to the test set label and actual output; Step 7: Calculate the gradient of the meta-learning loss function and use this gradient to update the residual neural network parameters; Step 8: Repeat steps 4 to 7 above for multiple times until the extraction of N subtask data is completed; Step 9: Input the communication modulation signal with the same properties as the final test signal into the residual neural network, repeat step 5, and complete the parameter update of the residual neural network; Step 10: Input the final test signal into the residual neural network to obtain the final classification result.

2. The communication signal modulation recognition method based on residual neural network and meta-learning fusion according to claim 1 is characterized in that: In step 1, the expressions of communication modulation signals with different signal-to-noise ratios under additive white Gaussian noise are: Wherein, s[t] and r[t] represent the transmission signal and the received signal in the tth time period, respectively, γ[t] is the communication channel gain, ω0 is the carrier frequency offset, θ0 is the phase offset between the sender and the receiver, v[t] is the additive white Gaussian noise, and j is the imaginary unit; The expression of signal-to-noise ratio is: Among them, P r is the power of the signal, P v is the power of the noise.

3. The communication signal modulation recognition method based on residual neural network and meta-learning fusion according to claim 1 is characterized in that: In step 2, the number of noise signal-to-noise ratios is N, the number of subtasks is N, each subtask data contains r types of communication signals, the types of communication signals contained in each subtask data are the same, the number of communication signal types contained in each subtask is the same as the number of communication modulation signal types used in the final test, and the number of samples in each subtask is the same.

4. The communication signal modulation recognition method based on residual neural network and meta-learning fusion according to claim 1 is characterized in that: In step 3, the residual neural network consists of 6 residual blocks, 2 fully connected layers, 1 Dropout layer and 1 Softmax layer, and the fully connected layer integrates the automatically extracted high-level features into local feature information; The Softmax layer is used for multi-class classification and its expression is: Among them, a k' is the k'th input signal of the output layer, y k' is the output of the k'th neuron, C is the number of output nodes, that is, the number of classification categories, a c Represents the cth input signal of the output layer.

5. The communication signal modulation recognition method based on residual neural network and meta-learning fusion according to claim 4 is characterized in that: Each residual block adopts the network structure of residual block 1 or residual block 2. The residual block 1 and residual block 2 are composed of a convolution layer, two sub-residual blocks and a pooling layer. Each sub-residual block contains 2 convolution layers. All convolution layers contain 32 convolution kernels, with a step size of 1 and a padding of 1. The convolution kernel sizes of the sub-residual blocks of residual block 1 and residual block 2 are different. The convolution kernel size of the sub-residual block of residual block 1 is 3×2, and the convolution kernel size of the sub-residual block of residual block 2 is 3×1. The convolution layer outside the sub-residual block is activated by the Relu function. The output of each sub-residual block is also activated by the Relu function and maximum pooling is performed. The convolution layer performs a convolution operation on the input to extract the characteristics of the signal. The expression of the convolution operation is: x l′ =f(w l′ *x l′-1 +b l′ ) Among them, l' represents the number of network layers, including input layer and convolution layer, w l′ represents the convolution kernel parameters in the current convolution layer, b l′ is the bias vector in the current convolutional layer, f(·) represents the activation function, x l′-1 Represents the output of the previous network layer, x l′ Represents the output of the current convolutional layer; The pooling layer downsamples the convolutional layer. The expression of the pooling layer is: x l′ =down(x l′-1 ) Where down(·) represents the pooling function; x l′-1 represents the output of the previous convolutional layer, x l′ Represents the output of the current pooling layer; The Relu function expression is: Among them, f(x) is the function that Relu takes the maximum value. When the input is negative, the output is 0, the neuron will not be activated, and x is the input of the function.

6. The communication signal modulation recognition method based on residual neural network and meta-learning fusion according to claim 1 is characterized in that: In step 4, the number of samples randomly selected from the subtask data each time is h+m, and the samples contain r types of communication signals, of which h samples are used as the training set of the extracted samples, and m samples are used as the test set of the extracted samples; the input data of the residual neural network is n in ×l×dim×1,n in is the number of input samples of the residual neural network, l is the signal sampling length, the signal is in complex form, the real part and the imaginary part of the signal are superimposed together, so dim=2.

7. The communication signal modulation recognition method based on residual neural network and meta-learning fusion according to claim 1 is characterized in that: In step 5, the steps to update the residual neural network parameters are: 1) Randomly initialize the residual neural network parameters φ0; 2) Input the extracted sample training set into the residual neural network; 3) Define the cross entropy loss function l(φ) using the extracted sample labels and network output results; Find the gradient of this function and update the parameters. Among them, k is the number of samples drawn, α is the inner loop learning rate, φ k is the residual neural network parameter when the kth sample is extracted, l(φ k ) is the cross entropy loss function of the network parameters, is the residual neural network parameter after gradient update; The cross entropy loss function is defined as: l(φ)=[y (k) logg(x (k) )+(1-y (k) )log(1-g(x (k) ))] Among them, x (k) is the data of the sample training set extracted for the kth time, y (k) is the label of the data of the k-th sample training set, g represents the residual neural network structure, and l(φ) is the cross entropy loss function.

8. The communication signal modulation recognition method based on residual neural network and meta-learning fusion according to claim 1 is characterized in that: In step 6, the expression of the meta-learning loss function is updated according to the labels corresponding to the test set and the output data corresponding to the test set: in, is the data of the sample test set extracted for the kth time, is the label of the data of the sample test set extracted for the kth time, is the meta-learning loss function updated after k-1 samples are drawn, 9. The communication signal modulation recognition method based on residual neural network and meta-learning fusion according to claim 1 is characterized in that: In step 7, the gradient of the meta-learning loss function is calculated, and the initial parameters of the residual neural network are updated using this gradient; the expression for updating the initial parameters of the residual neural network using the gradient is: Among them, β is the outer loop learning rate, and the inner loop learning rate α is greater than the outer loop learning rate β; is the updated initial parameter of the residual neural network.

10. The communication signal modulation recognition method based on residual neural network and meta-learning fusion according to claim 1 is characterized in that: In step 8, when repeating step 5, the parameters of the residual neural network are no longer randomly initialized, but are initialized to the network initial parameters updated in the previous iteration. In step 9, when repeating step 5, the parameters of the residual neural network are no longer randomly initialized, but are initialized to the initial parameters of the network after the last iteration update.

Images