An intelligent classification method for electromagnetic signal features based on lightweight progressively differentiable architecture search

Abstract

The present invention discloses an electromagnetic signal feature intelligent classification method based on lightweight progressively differentiable architecture search, comprising the following steps: step 1, using the pre-processed IQ samples generated by MATLAB as a proxy data set; step 2, constructing a search space and a unit structure to search for normal units and reduced units; step 3, searching for the best network architecture using a progressively differentiable architecture search method based on step 2; step 4, the training process of the final architecture; step 5, migrating the searched structure to the target task RadioML2016.10B; step 6, retraining to obtain the final model. The method disclosed by the present invention considers the classification accuracy and computational complexity of the model at the same time, achieves a balance between performance and computational overhead, and reduces the complexity of the model without almost reducing the modulation classification accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of automatic modulation classification and architecture search, and particularly relates to an electromagnetic signal feature intelligent classification method and model based on lightweight progressively differentiable architecture search in this field. Background Art

[0002] Automatic modulation classification is a technology for identifying the modulation type of received signals, including spontaneously classifying the modulation type of received signals under non-cooperative communication. It is widely used in various military and civilian scenarios. In military applications, it is mainly used to extract transmission information from encrypted modulated signals, and can also identify transmitter units in electronic warfare to generate matching interference signals. In modern civilian fields, it is often used to optimize transmission reliability and data rate.

[0003] Traditional automatic modulation classification methods mainly include maximum likelihood-based methods and feature-based methods. Maximum likelihood-based methods have high computational complexity and are sensitive to model mismatch and parameter deviation, while feature-based methods are highly dependent on the number of effective features. In recent years, deep learning methods have achieved great success in automatic modulation classification tasks. They rely on independent learning to analyze data rules, thereby getting rid of the dependence on artificial feature engineering and directly learning the mapping from raw signal data to actual output in an end-to-end manner. However, automatic modulation classification methods based on deep learning rely heavily on expert experience to design neural network structures, and lack flexibility due to their fixed architectures, which often leads to insufficient model generalization. Summary of the invention

[0004] In order to solve the problems that the existing traditional automatic modulation classification methods have high computational complexity and are highly dependent on the number of effective features, as well as the lack of flexibility and insufficient model generalization of the existing automatic modulation classification methods based on deep learning, the present invention provides an automatic modulation classification method and model based on lightweight progressively differentiable architecture search.

[0005] The present invention adopts the following technical solution:

[0006] An electromagnetic signal feature intelligent classification method based on lightweight progressively differentiable architecture search, the improvement of which is that it includes the following steps:

[0007] Step 1: Use the preprocessed IQ samples generated by MATLAB as a proxy dataset:

[0008] Set the sampling frequency to 8kHz, the carrier frequency to 2GHz, and the maximum Doppler shift to 100Hz; set the channel environment to Rayleigh channel, the signal-to-noise ratio range to 0dB~20dB, and the interval to 2dB;

[0009] The MATLAB candidate modulation pool has eight modulation signals, including three phase shift keying, three frequency shift keying, minimum frequency shift keying, and quadrature amplitude modulation. Each modulation type has 2,000 samples for training data and 1,000 samples for the test set of each signal-to-noise ratio, totaling 264,000 samples. In addition, the training data is divided into a training set and a validation set in a ratio of 7:3 for subsequent architecture search and evaluation.

[0010] The received baseband signal sequence r(n) is written as:

[0011] r(n)=he jθ(n) s(n)+w(n),n=1,2,…,N

[0012] In the above formula, s(n) represents the simulated signal sequence of the transmitter; w(n) is a sequence with a mean of 0 and a variance of is the additive white Gaussian noise; h is the fading channel gain; θ(n) represents the carrier phase offset;

[0013] The signal-to-noise ratio of the receiver is expressed as:

[0014]

[0015] According to the signal model, the received baseband signal sequence r(n) is decomposed into real and imaginary parts, which are written as:

[0016]

[0017]

[0018]

[0019] In the above formula, is the real part of r(n), is the imaginary part of r(n), and the IQ sample is given by and composition;

[0020] Step 2, construct the search space and cell structure to search for normal cells and reduced cells:

[0021] For normal cells, the input and output have the same size; for reduced cells, the size of the output is reduced by half;

[0022] A cell is a directed acyclic graph consisting of C nodes, where each node x i is a hidden representation. A unit c_{k} consists of two input nodes x0, x1 and an output node x C-1 and intermediate nodes x2,…,x C-2The input nodes come from the output nodes of the first two cells; the output nodes are the concatenation of the intermediate nodes; the directed edges e i,j represents information flow, and x i and x j An operation between i,j O is the set of all possible candidate operations, and each intermediate node is based on the information flow of the previous node, x j =∑ i＜j O i,j (x i );

[0023] The number of nodes is set to 7; the set O of candidate operation space includes the following eight operations:

[0024] 3×3 dilated convolution, 5×5 dilated convolution, 3×3 separable convolution, 5×5 separable convolution, 3×3 average pooling, 3×3 max pooling, residual connection, and none;

[0025] Specific operations between nodes are determined by a set of architectural parameters {α i,j}The weighted softmax of all possible operations is expressed as:

[0026]

[0027] In the above formula, o(x) represents a specific operation, and the goal of architecture search is to learn a series of continuous architecture parameters α = {α i,j}, and finally get an optimal operation:

[0028] Step 3: Based on step 2, use the progressively differentiable architecture search method to search for the best network architecture:

[0029] Randomly initialize model weights w0 and architecture parameters α0; create i,j Mixed operation O to candidate operation space O i,j ;

[0030] Step 31, search space approximation:

[0031] The entire search process is divided into the initial search stage, the intermediate search stage and the final search stage. Each search stage t has O t Candidate operations,

[0032] After each search phase, according to a set of learned weights w t Score the optional operations of each directed edge; directed edge e i,j The new candidate operation set O on t Less than O t-1 ;

[0033] At the end of the final search phase, according to a set of learned architecture parameters α t The final unit structure is derived; in each unit, the input node does not participate in the selection of candidate operations, and finally the first two nodes are connected to each intermediate node, and the connection between the nodes is to retain the candidate operation with the largest weight

[0034] Step 32, search space regularization:

[0035] At the beginning of each search phase, the model weights w and architecture parameters α are reinitialized;

[0036] Step 33, gradient-based search strategy:

[0037] By minimizing the validation loss L using the gradient descent technique val (w,α);

[0038] When convergence is not achieved, the iterative architecture parameter α is gradient-dropped by executing:

[0039] ξ represents the learning rate, L train (w,α) represents the training loss;

[0040] Iterate the model weights w in descending order and execute:

[0041]

[0042] Thus, the optimal architecture parameter α is obtained * ;

[0043] Step 4, training process of the final architecture:

[0044] Load the optimal architecture parameter α * ; Use cross entropy loss to train the model; Use the loss to update the training parameter θ, and finally save the optimal parameter θ * ;

[0045] Step 5: Migrate the searched structure to the target task RadioML2016.10B;

[0046] Step 6: Retrain to obtain the final model.

[0047] Further, in step 1: the three phase shift keyings are 2PSK, QPSK and 8PSK, and the three frequency shift keyings are 2FSK, 4FSK and 8FSK.

[0048] Further, in step 33:

[0049] When epoch=1,…,E-15, E is the number of epochs, and only the model weight w is fine-tuned;

[0050] When epoch = E-16,…,E, the adaptive moment estimation optimizer Adam and stochastic gradient descent SGD jointly optimize the model weights w and architecture parameters α.

[0051] Furthermore, in the three search stages of steps 31, 32 and 33, the sizes of the candidate operation space |O| are 8, 5 and 3 respectively; Dropout is set to 0, 0.5 and 0.5 respectively; in each stage, the batch size is set to 128 to train the network, the learning rate of each search stage is set to 0.1, the epoch is set to 25, and the model weight w is adjusted only in the first 10 epochs; the architecture parameter α and the model weight w are optimized by Adam and SGD in the last 15 epochs;

[0052] In the search phase of step 33, only one operation is retained to obtain the final normal unit and reduced unit, and the architecture learning rate and architecture weight decay of the Adam optimizer are set to 6×10 -4 and 10 -3 ,The network architecture includes a normal unit and a reduction unit, and the initial number of channels is set to 16.

[0053] Furthermore, in step 5, the signal-to-noise ratio of the samples is uniformly distributed in the range of -20dB to 18dB with an interval of 2dB. RadioML2016.10B contains ten different modulation types: BPSK, QPSK, 8PSK, 16QAM, 64QAM, BFSK, continuous phase shift keying CPFSK, four-level pulse amplitude modulation 4PAM, wideband frequency modulation WBFM and double-sideband amplitude modulation DSB-AM. Each modulation type has 6000 samples, and each signal-to-noise ratio has a total of 1,200,000 samples; the classic 3:1:1 division is used for training, validation and test sets.

[0054] Furthermore, in step 6, during the evaluation phase, the neural network is trained by setting the batch size, weight decay, and learning rate to 128, 3×10, respectively. -4 and 0.025, with 32 initial channels and 100 epochs.

[0055] An electromagnetic signal feature intelligent classification model based on lightweight progressively differentiable architecture search, the improvement of which lies in: the model is implemented by the above method steps.

[0056] The beneficial effects of the present invention are:

[0057] The method disclosed in the present invention takes into account the classification accuracy and computational complexity of the model at the same time, achieves a balance between performance and computational overhead, and reduces the complexity of the model without reducing the modulation classification accuracy. At the same time, the optimal architecture searched on the MATLAB simulated data set was migrated to the RadioML2016.10B task, verifying the robustness and generalization ability of the proposed method, which can greatly promote the productization of this field and provide an effective solution for the construction of automatic modulation signal classification model.

[0058] The method disclosed in the present invention transfers the neural architecture search to the field of automatic modulation classification, and can find the optimal neural network architecture in an automated way, thereby improving the model performance and efficiency, and is more conducive to industrial mass production. At the same time, the proposed model has good portability, scalability and sustainability, and has great practical significance in practical applications.

[0059] The method disclosed in the present invention solves the problem of high computational complexity of traditional automatic modulation classification methods, effectively achieves a balance between computational overhead and classification performance, and also solves the problems of poor transferability and insufficient generalization ability of automatic modulation classification methods based on deep learning, which is helpful for subsequent productization and large-scale production. It can not only ensure the classification accuracy and robustness of the model, but also flexibly design the network structure, thereby enhancing the flexibility, scalability and sustainability of the model.

[0060] The model disclosed in the present invention can use the PDARTS-AMC algorithm designed in step 2 and step 3 to search for the best normal unit and reduced unit, and can quickly and effectively optimize the units that constitute the neural network architecture through a search algorithm based on progressive differentiability; the predefined search space finds the candidate network structure through the search strategy, evaluates the network architecture, and conducts the next round of search based on the feedback of the performance evaluation; in step 4, the searched units are stacked according to the target task, and a separable convolution operation is used to create a lightweight network to reduce the complexity of the network model, which solves the problem of high computational complexity of traditional automatic modulation classification methods and effectively achieves a balance between computational overhead and classification performance; through step 5, the searched structure is migrated to the target task RadioML2016.10B to prove the robustness and generalization ability of the model, which solves the problems of lack of flexibility and insufficient model generalization of the automatic modulation classification method based on deep learning, and further enhances the transferability, scalability and sustainability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic flow diagram of the method of the present invention;

[0062] Figure 2 It is a schematic diagram of the comparison of the architecture accuracy of different unit stacking of the model of the present invention in the MATLAB data set;

[0063] Figure 3 It is a schematic diagram of the comparison of the architecture accuracy of different unit stacking of the model of the present invention in the RadioML2016.10B data set;

[0064] Figure 4 It is a schematic diagram comparing the search performance and complexity of different methods of the model of the present invention in the MATLAB data set. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0066] Neural architecture search is a new attempt to introduce automatic modulation classification. It can automatically design deep neural network models based on received signal samples without requiring a lot of expertise and repeated experiments. It can automatically generate network structures that are comparable to or superior to manually designed classic networks (such as ResNet, VGG, etc.), with fewer parameters and faster speed. For example, NASNet, PNASNet, AmoebaNet, etc. are all excellent classification networks obtained through neural architecture search.

[0067] Embodiment 1, this embodiment discloses an electromagnetic signal feature intelligent classification method based on lightweight progressively differentiable architecture search, such as Figure 1 As shown, the following steps are included:

[0068] Step 1: Use the preprocessed IQ samples generated by MATLAB as a proxy dataset:

[0069] Set the sampling frequency to 8kHz, the carrier frequency to 2GHz, and the maximum Doppler shift to 100Hz; set the channel environment to Rayleigh channel, the signal-to-noise ratio range to 0dB~20dB, and the interval to 2dB;

[0070] The MATLAB candidate modulation pool has eight modulation signals, including three phase shift keying (2PSK, QPSK, 8PSK), three frequency shift keying (2FSK, 4FSK, 8FSK), minimum frequency shift keying (MSK) and quadrature amplitude modulation (QAM). Each modulation type has 2000 samples for training data and 1000 samples for the test set of each signal-to-noise ratio, totaling 264000 samples. In addition, the training data is divided into training set and validation set in a ratio of 7:3 for subsequent architecture search and evaluation.

[0071] The received baseband signal sequence r(n) is written as:

[0072] r(n)=he jθ(n) s(n)+w(n),n=1,2,…,N

[0073] In the above formula, s(n) represents the simulated signal sequence of the transmitter; w(n) is a sequence with a mean of 0 and a variance of is the additive white Gaussian noise; h is the fading channel gain; θ(n) represents the carrier phase offset;

[0074] The signal-to-noise ratio of the receiver is expressed as:

[0075]

[0076] According to the signal model, the received baseband signal sequence r(n) is decomposed into real and imaginary parts, which are written as:

[0077]

[0078]

[0079]

[0080] In the above formula, is the real part of r(n), is the imaginary part of r(n), and the IQ sample is given by and It can be used as an input dataset in a neural network.

[0081] Step 2, construct the search space and cell structure to search for normal cells and reduced cells:

[0082] The search space and unit structure are introduced. The essence is to search the network architecture based on the unit (convolution, pooling, residual connection, etc.). The complete network architecture is composed of the searched units.

[0083] This step mainly searches for two types of units, normal units and reduced units. For normal units, the input and output have the same size; for reduced units, the size of the output is reduced by half;

[0084] A cell is a directed acyclic graph consisting of C nodes, where each node x i is a hidden representation. A unit c_{k} consists of two input nodes x0, x1 and an output node x C-1 and intermediate nodes x2,…,x C-2 The input nodes come from the output nodes of the first two cells (i.e., c_{k-2} and c_{k-1}); the output node is the concatenation of the intermediate nodes, i.e., x C-1 =concat(x2,…,x C-2 ); directed edge ei,j represents information flow, and x i and x j An operation between i,j O is the set of all possible candidate operations (such as convolution, pooling, and residual connection), and each intermediate node is based on the information flow of the previous node, x j =∑ i＜j O i,j (x i );

[0085] The number of nodes is set to 7; the set O of candidate operation space includes the following eight operations:

[0086] 3×3 dilated convolution, 5×5 dilated convolution, 3×3 separable convolution, 5×5 separable convolution, 3×3 average pooling, 3×3 max pooling, residual connection, and none; dilated convolution increases the receptive field while maintaining the size of the feature map. Compared with traditional convolution operations, separable convolution has a lower number of parameters and operation cost. Residual connection solves the problems of gradient explosion and gradient disappearance during training. Pooling operation retains the main features while compressing them to simplify the network complexity.

[0087] In order to transform the discrete search space into a continuous one, the specific operations between nodes are relaxed to be defined by a set of architectural parameters {α i,j}The weighted softmax of all possible operations is expressed as:

[0088]

[0089] In the above formula, o(x) represents a specific computing operation, such as a convolution operation or a skip connection. The goal of architecture search is to learn a series of continuous architecture parameters α = {α i,j}, at the end of each architecture search phase, the connections between nodes are gradually replaced by mixed operations with several most likely operations, and finally an optimal operation is obtained:

[0090] Step 3: Based on step 2, use the progressively differentiable architecture search method to search for the best network architecture:

[0091] Randomly initialize model weights w0 and architecture parameters α0; create i,j Mixed operation O to candidate operation space O i,j ;

[0092] Step 31, search space approximation:

[0093] The entire search process is divided into the initial search stage, the intermediate search stage and the final search stage. Each search stage t has O t Candidate operations,

[0094] After each search phase, according to a set of learned weights w t The optional operations of each directed edge are scored; the approximation criterion means that candidate operations with lower weights are removed in the previous search phase, so the directed edge e i,j The new candidate operation set O on t Less than O t-1 ;

[0095] At the end of the final search phase, according to a set of learned architecture parameters α t The final unit structure is derived; in each unit, the input node does not participate in the selection of candidate operations, and finally the first two nodes are connected to each intermediate node, and the connection between the nodes is to retain the candidate operation with the largest weight

[0096] Step 32, search space regularization:

[0097] At the beginning of each search phase, the model weights w and architecture parameters α are reinitialized; that is, the network is trained from scratch.

[0098] On small datasets that are relatively easy to fit, residual connection operations are more likely to be selected than other candidate operations such as convolution and pooling. Since the gradient of residual connection with respect to the input is always 1, it can speed up forward and backward propagation. The gradient of convolution operation is much smaller than residual connection, and its output has no significant weight in classification. The operations with larger weights in the unit are retained, and the weights of other operations are cleared to 0.

[0099] In order to solve the problem of residual connection enrichment, the search space is regularized through architectural optimization, that is, the number of residual connection operations that are finally retained in the unit is limited through an iterative process. If the number of residual connection operations is insufficient, only the operations with larger weights in the unit are retained to reconstruct the unit structure.

[0100] Step 33, gradient-based search strategy:

[0101] By minimizing the validation loss L using the gradient descent technique val (w,α); L val (w,α) is closely related to the architecture parameters α and the network weights w, so the goal is to jointly learn the architecture parameters α and the network weights w of all hybrid candidate operations, where w is similar to the weights of the fully connected layers.

[0102] The goal of architecture search is to find the minimum L val (w * ,α * ) in the best architecture α * , where w* is obtained by minimizing the training loss, i.e., w * = arg min w L train (w,α * ). This becomes a two-level optimization problem, which can be written as:

[0103]

[0104]

[0105] The architecture α can be viewed as a special type of hyperparameter that can be optimized using gradient-based optimization methods and can be expressed as:

[0106]

[0107] Where ξ is the learning rate. The main idea is to use the training step w * (α) to approximate Bring the model to convergence, rather than just optimizing internally.

[0108] The specific process is:

[0109] When convergence is not achieved, the iterative architecture parameter α is gradient-dropped by executing:

[0110] ξ represents the learning rate, L train (w,α) represents the training loss;

[0111] Iterate the model weights w in descending order and execute:

[0112]

[0113] When epoch=1,…,E-15, E is the number of epochs, and only the model weight w is fine-tuned;

[0114] When epoch = E-16,…,E, the Adaptive Moment Estimation optimizer (Adam) and Stochastic Gradient Descent (SGD) jointly optimize the model weights w and architecture parameters α.

[0115] Thus, the optimal architecture parameter α is obtained * ;

[0116] In the three search stages of steps 31, 32 and 33, the sizes of the candidate operation space |O| are 8, 5 and 3 respectively; Dropout is set to 0, 0.5 and 0.5 respectively; in each stage, the batch size is set to 128 to train the network, the learning rate of each search stage is set to 0.1, the epoch is set to 25, and the model weight w is only adjusted in the first 10 epochs; the architecture parameter α and the model weight w are optimized by Adam and SGD in the last 15 epochs;

[0117] In the search phase of step 33, only one operation is retained to obtain the final normal unit and reduced unit, and the architecture learning rate and architecture weight decay of the Adam optimizer are set to 6×10 -4 and 10 -3 ,The network architecture includes a normal unit and a reduction unit, and the initial number of channels is set to 16.

[0118] Step 4, training process of the final architecture:

[0119] The searched units are stacked according to the target task to form network architectures of different depths. The final network architecture is used to extract the features of the received modulated signal samples under different signal-to-noise ratio environments. At the same time, separable convolution operations are used to reduce the complexity of the network model, and pooling operations are used to compress distinguishable signal features. The cross entropy loss function is used to describe the distance between the true output and the expected output.

[0120] The training process of the final architecture is as follows:

[0121] Load the optimal architecture parameter α * ; Use cross entropy loss to train the model; Use the loss to update the training parameter θ, and finally save the optimal parameter θ * ;

[0122] Step 5, transfer the searched structure to the target task RadioML2016.10B, which proves the robustness and generalization ability of the present invention:

[0123] The RadioML2016.10B dataset is generated by GNU Radio, which is corrupted by AWGN and has selective fading (including Rician and Rayleigh fading), sampling rate offset, and center frequency offset, which can reflect various defects of real wireless channels.

[0124] The signal-to-noise ratio of the samples is evenly distributed in the range of -20dB to 18dB with an interval of 2dB. RadioML2016.10B contains ten different modulation types: BPSK, QPSK, 8PSK, 16QAM, 64QAM, BFSK, continuous phase-shift keying CPFSK, four-level pulse amplitude modulation 4PAM, wideband frequency modulation WBFM and double-sideband amplitude modulation DSB-AM. Each modulation type has 6000 samples, and each signal-to-noise ratio has a total of 1,200,000 samples. The classic 3:1:1 division is used for training, validation and test sets.

[0125] Step 6: Retrain to obtain the final model.

[0126] In the evaluation phase, the neural network is trained by setting the batch size, weight decay, and learning rate to 128, 3×10, respectively. -4 and 0.025, with an initial channel of 32 and an epoch of 100. The final architecture is trained from scratch to achieve high modulation classification accuracy, and high-performance modulation classification results are obtained.

[0127] like Figure 4 The figure shows the comparison between the proposed method and different architecture search methods. The proposed method only takes 0.1632 GPU days on the MATLAB simulated data set, and the test error is 0.0692. The search space approximation technique reduces the search cost to 0.1632 GPU days, which shows that the search space approximation scheme of the proposed method has certain efficiency performance. The parameter and test errors of the proposed method are slightly smaller than those of DARTS and NAS, but the time spent searching the final architecture is much less.

[0128] This embodiment also discloses an electromagnetic signal feature intelligent classification model based on lightweight progressively differentiable architecture search, which is implemented by the above method steps.

[0129] After three search stages, the optimal normal units and reduced units can be obtained, the best stacking can be used to form the final network architecture, and the architecture can be trained from scratch to obtain higher classification accuracy under different signal-to-noise ratio environments. Figure 2 As shown in Figure 1, on the dataset simulated in MATLAB, the modulation classification accuracy of the network architectures stacked by different numbers of units is roughly the same. Therefore, the architecture stacked by one normal unit and one reduced unit can fully identify the modulation type of the received signal.

[0130] The architecture searched on the MATLAB simulated dataset is deployed to more realistic communication scenarios, which requires it to be highly general and robust. Therefore, the architecture searched on the MATLAB simulated dataset is transferred to RadioML2016.10B. Figure 3 It can be seen that the test accuracy is the lowest when the number of stacked units is 2. This is mainly because RadioML2016.10B has much more data than the dataset simulated by MATLAB, and the network architecture is shallow and cannot learn the modulated signal features at various abstract levels. When the number of units is 6, the average accuracy is the highest. However, when the number of units is 8, the classification accuracy decreases. This is mainly because over-deep neural networks can bear the risk of overfitting. Therefore, for RadioML2016.10B, 6 units are stacked to fully recognize the modulated signal.

Claims

1. An electromagnetic signal feature intelligent classification method based on lightweight progressively differentiable architecture search, characterized in that: The steps include: Step 1: Use the preprocessed IQ samples generated by MATLAB as a proxy dataset: Set the sampling frequency to 8kHz, the carrier frequency to 2GHz, and the maximum Doppler shift to 100Hz; set the channel environment to Rayleigh channel, the signal-to-noise ratio range to 0dB~20dB, and the interval to 2dB; The MATLAB candidate modulation pool has eight modulation signals, including three phase shift keying, three frequency shift keying, minimum frequency shift keying, and quadrature amplitude modulation. Each modulation type has 2,000 samples for training data and 1,000 samples for the test set of each signal-to-noise ratio, totaling 264,000 samples. In addition, the training data is divided into a training set and a validation set in a ratio of 7:3 for subsequent architecture search and evaluation. The received baseband signal sequence r(n) is written as: r(n)=he jθ(n) s(n)+w(n),n=1,2,…,N In the above formula, s(n) represents the simulated signal sequence of the transmitter; w(n) is a sequence with a mean of 0 and a variance of is the additive white Gaussian noise; h is the fading channel gain; θ(n) represents the carrier phase offset; The signal-to-noise ratio of the receiver is expressed as: According to the signal model, the received baseband signal sequence r(n) is decomposed into real and imaginary parts, which are written as: In the above formula, is the real part of r(n), is the imaginary part of r(n), and the IQ sample is given by and composition; Step 2, construct the search space and cell structure to search for normal cells and reduced cells: For normal cells, the input and output have the same size; for reduced cells, the size of the output is reduced by half; A cell is a directed acyclic graph consisting of C nodes, where each node x i is a hidden representation. A unit c_{k} consists of two input nodes x0, x1 and an output node x C-1 and intermediate nodes x2,…,x C-2 The input nodes come from the output nodes of the first two cells; the output nodes are the concatenation of the intermediate nodes; the directed edges e i,j represents information flow, and x i and x j An operation between i,j O is the set of all possible candidate operations, and each intermediate node is based on the information flow of the previous node, x j =∑ i＜j O i,j (x i ); The number of nodes is set to 7; the set O of candidate operation space includes the following eight operations: 3×3 dilated convolution, 5×5 dilated convolution, 3×3 separable convolution, 5×5 separable convolution, 3×3 average pooling, 3×3 max pooling, residual connection, and none; Specific operations between nodes are determined by a set of architectural parameters {α i,j }The weighted softmax of all possible operations is expressed as: In the above formula, o(x) represents a specific operation, and the goal of architecture search is to learn a series of continuous architecture parameters α = {α i,j }, and finally get an optimal operation: Step 3: Based on step 2, use the progressively differentiable architecture search method to search for the best network architecture: Randomly initialize the model weights w0 and architecture parameters α0; create i,j Mixed operation O to candidate operation space O i,j ; Step 31, search space approximation: The entire search process is divided into the initial search stage, the intermediate search stage and the final search stage. Each search stage t has O t Candidate operations, After each search phase, according to a set of learned weights w t Score the optional operations of each directed edge; directed edge e i,j The new candidate operation set O on t Less than O t-1 ; At the end of the final search phase, according to a set of learned architecture parameters α t The final unit structure is derived; in each unit, the input node does not participate in the selection of candidate operations, and finally the first two nodes are connected to each intermediate node, and the connection between the nodes is to retain the candidate operation with the largest weight Step 32, search space regularization: At the beginning of each search phase, the model weights w and architecture parameters α are reinitialized; Step 33, gradient-based search strategy: By minimizing the validation loss L using the gradient descent technique val (w,α); When convergence is not achieved, the iterative architecture parameter α is gradient-dropped by executing: ▽ α L val (w-ξ▽ w L train (w,α),α), ξ represents the learning rate, L train (w,α) represents the training loss; Iterate the model weights w in descending order and execute: ▽ w L train (w,a); Thus, the optimal architecture parameter α is obtained * ; Step 4, training process of the final architecture: Load the optimal architecture parameter α * ; Use cross entropy loss to train the model; Use the loss to update the training parameter θ, and finally save the optimal parameter θ * ; Step 5: Migrate the searched structure to the target task RadioML2016.10B; Step 6: Retrain to obtain the final model.

2. The electromagnetic signal feature intelligent classification method based on lightweight progressively differentiable architecture search according to claim 1 is characterized by: In step 1: the three phase shift keyings are 2PSK, QPSK and 8PSK, and the three frequency shift keyings are 2FSK, 4FSK and 8FSK.

3. The electromagnetic signal feature intelligent classification method based on lightweight progressively differentiable architecture search according to claim 1 is characterized by: In step 33: When epoch=1,…,E-15, E is the number of epochs, and only the model weight w is fine-tuned; When epoch = E-16,…,E, the adaptive moment estimation optimizer Adam and stochastic gradient descent SGD jointly optimize the model weights w and architecture parameters α.

4. The electromagnetic signal feature intelligent classification method based on lightweight progressively differentiable architecture search according to claim 1 is characterized by: In the three search stages of steps 31, 32 and 33, the sizes of the candidate operation space |O| are 8, 5 and 3 respectively; Dropout is set to 0, 0.5 and 0.5 respectively; in each stage, the batch size is set to 128 to train the network, the learning rate of each search stage is set to 0.1, the epoch is set to 25, and the model weight w is only adjusted in the first 10 epochs; the architecture parameter α and the model weight w are optimized by Adam and SGD in the last 15 epochs; In the search phase of step 33, only one operation is retained to obtain the final normal unit and reduced unit, and the architecture learning rate and architecture weight decay of the Adam optimizer are set to 6×10 -4 and 10 -3 ,The network architecture includes a normal unit and a reduction unit, and the initial number of channels is set to 16.

5. The electromagnetic signal feature intelligent classification method based on lightweight progressively differentiable architecture search according to claim 1 is characterized by: In step 5, the signal-to-noise ratio of the samples is evenly distributed in the range of -20dB to 18dB with an interval of 2dB. RadioML2016.10B contains ten different modulation types: BPSK, QPSK, 8PSK, 16QAM, 64QAM, BFSK, continuous phase shift keying CPFSK, four-level pulse amplitude modulation 4PAM, wideband frequency modulation WBFM and double-sideband amplitude modulation DSB-AM. Each modulation type has 6000 samples, and each signal-to-noise ratio has a total of 1,200,000 samples; the classic 3:1:1 division is used for training, validation and test sets.

6. The electromagnetic signal feature intelligent classification method based on lightweight progressively differentiable architecture search according to claim 1 is characterized by: Step 6 In the evaluation phase, the neural network is evaluated by setting the batch size, weight decay, and learning rate to 128, 3×10, respectively. -4 and 0.025, with 32 initial channels and 100 epochs.

Images