Federal learning modulation identification method and device for heterogeneous data

Abstract

The invention discloses a federal learning modulation identification method and device for heterogeneous data, and relates to the technical field of radio modulation signal automatic identification. According to the method, radio modulation signals are received, signal acquisition and preprocessing are performed on the received signals, preprocessed feature vectors are input into a federated learning network, the rth round of federated learning is started, and after a round of local training is completed by a local model of a kth computing node in the network, the rth round of federated learning is completed. And obtaining the weight of the updated local model of the kth computing node, uploading the weights of the local models of the computing nodes subjected to pruning to a server through an uplink for aggregation so as to obtain the weight of a global model, thereby identifying the classification of the radio modulation signals. The method is oriented to heterogeneous data, diversified data from different clients can be effectively integrated, and the recognition accuracy and robustness of the model to different modulation types are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of radio signal modulation recognition, and in particular to a method and device for automatically recognizing the modulation of a radio communication signal through a federated learning network. Background Art

[0002] In radio communication systems in complex electromagnetic environments, radio communication signals are usually modulated. The purpose is mainly to determine the modulation mode and some other signal parameters of the received signal in a complex signal environment. With the Internet access of edge client devices in communication systems, the demand for efficient data transmission of edge client devices has been proposed. At the same time, with the massive growth of edge client devices, the demand for efficient data transmission is in conflict with signal and spectrum management. Signal modulation recognition technology can realize signal and spectrum management. Without sufficient prior knowledge, it can analyze and process the received signal and finally determine the modulation mode of the received signal. Therefore, it plays an increasingly important role in ensuring efficient data transmission. In recent years, the successful application of deep learning in computer vision, natural language and other fields has also brought more in-depth exploration to signal modulation recognition technology. Automatic signal modulation recognition methods based on deep learning have been successfully applied to classify the modulation types of signals in complex electromagnetic environments. Among them, most researchers have improved the classification and recognition performance by using centralized deep learning methods. This centralized deep learning method requires each edge client device to send data to a central server, complete data storage and computation processing on the central server, and use complex data input into a large network model to improve recognition performance. However, the above methods require complex data and network models to be stored and calculated on a central server, which places high demands on the data storage and computing capabilities of the central server, making it difficult to adapt to the rapid growth of massive data volumes and the number of edge client devices. Therefore, distributed learning has begun to be used to improve the performance of signal modulation recognition technology. Using multiple edge client devices as distributed nodes during data storage and computing can alleviate the difficulties faced by the above centralized deep learning.

[0003] As a special distributed learning method that uses data parallelism, federated learning can emphasize the local data differences of multiple edge client devices. It does not need to store and calculate complex data on a central server, and can protect the client's data privacy to a greater extent. At the same time, distributed training is performed in a data parallel manner to converge to a high-performance global model, which is more conducive to the dynamic access and perception of each edge client device in the communication network. At present, the use of federated learning for modulation recognition is often limited by the data quality and diversity of each computing node. If the data distribution of each computing node does not conform to independent and identical distribution, it will cause deviations in the model training results and affect the final recognition accuracy. In addition, how to improve computing efficiency is also an important issue that needs to be solved when applying federated learning for adaptive modulation recognition. In the process of model upload aggregation, a large communication bandwidth is required, which will also increase the computing delay. Especially when processing large-scale data, this delay will be more significant. Therefore, it is necessary to explore new federated learning strategies to improve training effects and computing efficiency.

[0004] The present invention aims to solve a series of challenges faced by existing radio signal modulation recognition in a federated learning environment. In particular, in a federated architecture, the local data sets of each computing node have the characteristics of non-independent and identically distributed (Non-IID), which often leads to a significant decrease in the overall recognition performance. In addition, the existing technology also has deficiencies in communication and computing efficiency, which are particularly prominent when processing large-scale data sets.

[0005] From the perspective of practical application, the present invention proposes a federated modulation recognition method for heterogeneous data. The method adopts adaptive model pruning to solve the challenge of non-independent and identically distributed data. The model complexity can be adaptively adjusted based on the characteristics of the local data set of each computing node, thereby improving the training effect and computing efficiency of the model and making the information link transmission more efficient. Summary of the invention

[0006] The present invention proposes an efficient federated modulation recognition technology under the condition that the server and the client adopt a synchronous communication mode. The present invention receives a radio modulation signal, performs signal acquisition and preprocessing on the received signal, inputs the preprocessed feature vector into the federated learning network, and starts the rth round of federated learning. After completing a round of local training, the local model of the kth computing node in the network obtains the weight of the updated local model of the kth computing node, and uploads the weight of the pruned local model of each computing node to the server via the uplink for aggregation to obtain the weight of the global model, thereby identifying the classification of the radio modulation signal. This technology is oriented to heterogeneous data, can effectively integrate diversified data from different clients, improve the recognition accuracy and robustness of the model for different modulation types, and provide a solution for building a more intelligent, secure and reliable communication network.

[0007] In the first aspect, a federated learning modulation recognition method for heterogeneous data is disclosed, characterized in that it includes the following steps: S1: receiving a radio modulation signal, and performing signal acquisition and preprocessing on the received signal, and inputting the preprocessed feature vector into a federated learning network;

[0008] S2: The federated learning network includes k computing nodes, where k=1…K, K represents the number of computing nodes in the federated learning network. The rth round of federated learning begins. The kth computing node receives the weight w(r) of the global model in the rth round of communication. Then the weight of the local model of the kth computing node is w k (r), after completing r rounds of local training, the local model of the k-th computing node obtains the weight w of the updated local model of the k-th computing node k (r+1);

[0009] First, the updated weight of the kth computing node is w k The local model of (r+1) needs to calculate the similarity between the weights of each layer and the weights of the corresponding layer in the global model with weight w(r) It can be expressed as:

[0010]

[0011] in, represents the updated weight of the lth layer of node k after local training in the rth round of training, w l (r) represents the weight of the received global model layer l, where l is a positive integer from 1 to L, L represents the total number of layers in the network, and k and r are positive integers;

[0012] Next, calculate the pruning rate of the lth layer of node k Defined as:

[0013]

[0014] Among them, γ is the preset hyperparameter,

[0015] Finally, according to the obtained pruning rate Calculate the weight of the lth layer Threshold The mask matrix m is derived k (r+1), when the weight value of the lth layer When it is lower than the threshold, the corresponding value of the mask matrix needs to be set to 0, otherwise it is set to 1, which can be expressed as:

[0016]

[0017] The final mask matrix obtained is m k (r+1), when the local model is pruned, it is represented as Where ⊙ represents Hadamard convolution. After the pruning of the computing node k in the rth round of training, the weight of the local model is updated to w k (r+1);

[0018] S3: The weights of the local models of each computing node after pruning are uploaded to the server via the uplink. The server calculates the weights w of the received local models. k (r+1) is averaged and aggregated to obtain a global model. The weight of the global model obtained by the aggregation is expressed as w(r+1). The process can be defined as:

[0019]

[0020] After the rth round of federated learning is completed, steps S1, S2, and S3 are repeated to proceed to the next round until the performance changes of the local model and the global model tend to zero, and the global model feature vector and the local model feature vector of the feature vector of the serially connected radio modulation signal are used for signal modulation identification to obtain a modulation classification result; the global model feature vector at least includes a global model weight, and the local model feature vector at least includes a local model weight.

[0021] In the second aspect, the step S1 further comprises the following steps:

[0022] S11: receiving radio modulation signals for signal acquisition and preprocessing;

[0023] S12: Build a federated learning network model. The entire construction process goes through four parts: one-dimensional convolution, deep residual shrinkage network DSRN, GRU unit, and fully connected layer (Full-Connected Layer, FC).

[0024] In the third aspect, step S12 further comprises:

[0025] Through a layer of one-dimensional convolution Conv1d,(2,32) and Conv1d,(1,32) with input channel numbers of 2 and 1, output channel numbers of 32, convolution kernel sizes k of 3, and moving step sizes s of 2, the data features of the I / Q signal and the phase information of the P value are extracted, and two feature vectors with data sizes of 2×1024 and 1×1024 are obtained, wherein the I / Q signal is a baseband signal of a radio modulation signal obtained from signal acquisition and preprocessing, which is stored and processed in I / Q form, wherein I represents the in-phase component, Q represents the orthogonal component, and the P value signal represents the instantaneous phase of the baseband signal of the radio modulation signal; the two feature vectors are spliced ​​and integrated into a single feature vector, and the output data size is a feature vector of 32×1024;

[0026] Continue to pass the feature vector with a data size of 32×1024 through 5 stacked DRSN modules to output a feature vector with a data size of 32×32;

[0027] Next, the output from the DRSN module passes through the GRU unit so that the GRU unit captures the contextual relationship of the feature vector;

[0028] Finally, the feature vector output by the GRU unit is input into the fully connected layer FC, and then the probability of the feature vector belonging to each category is obtained through the Softmax activation function.

[0029] In the fourth aspect, through a layer of one-dimensional convolution Conv1d,(2,32) and Conv1d,(1,32) with 2 and 1 input channels, 32 output channels, 3 convolution kernel sizes k, and 2 moving steps s, the data features of the I / Q signal and the phase information of the P value are extracted, and two feature vectors with data sizes of 2×1024 and 1×1024 are obtained, wherein the I / Q signal is a baseband signal of a radio modulation signal obtained from signal acquisition and preprocessing, which is stored and processed in I / Q form, wherein I represents the in-phase component, Q represents the orthogonal component, and the P value signal represents the instantaneous phase of the baseband signal of the radio modulation signal; the two feature vectors are spliced ​​and integrated into a single feature vector, and the output data size is a feature vector of 32×1024;

[0030] Continue to pass the feature vector with a data size of 32×1024 through 5 stacked DRSN modules to output a feature vector with a data size of 32×32;

[0031] Next, the output from the DRSN module passes through the GRU unit so that the GRU unit captures the contextual relationship of the feature vector;

[0032] Finally, the feature vector output by the GRU unit is input into the fully connected layer FC, and then the probability of the feature vector belonging to each category is obtained through the Softmax activation function.

[0033] In a fifth aspect, a federated learning modulation recognition device for heterogeneous data is disclosed, characterized in that the device comprises the following:

[0034] An acquisition and preprocessing module, which receives a radio modulation signal, performs signal acquisition and preprocessing on the received signal, and inputs the preprocessed feature vector into the federated learning network;

[0035] A local model training module, wherein the local model training module performs training and updating of the local model, wherein the federated learning network includes k computing nodes, wherein k=1…K, K represents the number of computing nodes in the federated learning network, and the rth round of federated learning is started. The kth computing node receives the weight w(r) of the global model in the rth round of communication, and the weight of the local model of the kth computing node is w k (r), after completing r rounds of local training, the local model of the k-th computing node obtains the weight w of the updated local model of the k-th computing node k (r+1);

[0036] First, the updated weight of the kth computing node is w k The local model of (r+1) needs to calculate the similarity between the weights of each layer and the weights of the corresponding layer in the global model with weight w(r) It can be expressed as:

[0037]

[0038] in, represents the updated weight of the lth layer of node k after local training in the rth round of training, w l (r) represents the weight of the received global model layer l, where l is a positive integer from 1 to L, L represents the total number of layers in the network, and k and r are positive integers;

[0039] Next, calculate the pruning rate of the lth layer of node k Defined as:

[0040]

[0041] Among them, γ is the preset hyperparameter,

[0042] Finally, according to the obtained pruning rate Calculate the weight of the lth layer Threshold The mask matrix m is derived k (r+1), when the weight value of the lth layer When it is lower than the threshold, the corresponding value of the mask matrix needs to be set to 0, otherwise it is set to 1, which can be expressed as:

[0043]

[0044] The final mask matrix obtained is m k (r+1), when the local model is pruned, it is represented as Where ⊙ represents Hadamard convolution. After the pruning of the computing node k in the rth round of training, the weight of the local model is updated to w k (r+1);

[0045] The aggregation module uploads the weights of the local models of each computing node after pruning to the server via an uplink. The server calculates the weights of the local models received by the server. k (r+1) is averaged and aggregated to obtain a global model. The weight of the global model obtained by the aggregation is expressed as w(r+1). The process can be defined as:

[0046]

[0047] After the device performs the rth round of federated learning, the acquisition and preprocessing module, the local model training module and the aggregation module in the device repeat the next round of federated learning until the performance change of the local model and the global model tends to zero, and the global model feature vector and the local model feature vector of the feature vector of the serially connected radio modulation signal are used for signal modulation identification to obtain a modulation classification result; the global model feature vector includes at least a global model weight, and the local model feature vector includes at least a local model weight.

[0048] In a sixth aspect, the local model training module further includes an acquisition and preprocessing submodule and a federated learning network model building submodule;

[0049] Among them, the acquisition and preprocessing submodule is responsible for receiving radio modulation signals for signal acquisition and preprocessing;

[0050] The federated learning network model construction submodule is responsible for constructing the federated learning network model. The entire construction process goes through four parts: one-dimensional convolution, deep residual shrinkage network DSRN, GRU unit, and fully connected layer (Full-Connected Layer, FC).

[0051] In the seventh aspect, the federated learning network model construction submodule further includes the following units:

[0052] A layer of convolution units, the layer of convolution units is used to extract the data features of the I / Q signal and the phase information of the P value, through a layer of one-dimensional convolution Conv1d, (2, 32) and Conv1d, (1, 32) with input channel numbers of 2 and 1, output channel numbers of 32, convolution kernel sizes k of 3, and moving step sizes s of 2, to obtain two feature vectors with data sizes of 2×1024 and 1×1024 respectively, the two feature vectors are spliced ​​and integrated into a single feature vector, and the final output feature vector has a feature size of 32×1024;

[0053] The DRSN module unit continues to pass the feature vector obtained by the convolutional unit of the first layer through 5 stacked DRSN modules, so that the output feature size is a feature vector of 32×32;

[0054] GRU unit, the output from the DRSN module unit needs to pass through a GRU unit with an input size of 32 and a hidden layer number of 128, so that the GRU unit can capture the contextual relationship of the feature vector;

[0055] The fully connected layer FC unit maps the feature vector output by the GRU unit to the final output through a fully connected layer FC, the output size is 1×24, and the probability that the input feature vector belongs to each category is obtained through the Softmax activation function, and the category of the radio modulation signal is determined according to the probability.

[0056] In an eighth aspect, a federated learning modulation recognition device for heterogeneous data is disclosed, characterized in that the device comprises the following:

[0057] A local model training module receives a radio modulated signal for signal acquisition and preprocessing, and inputs the preprocessed feature vector into a federated learning network, wherein the federated learning network includes k distributed computing nodes, where k=1…K, K represents the number of computing nodes in the federated learning network, and starts the rth round of federated learning. The weight w(r) of the initial global model sent by the server in the network via the downlink is received, and the local model weight of the kth node is w k (r); local model w located at each of the computing nodes k (r) The neural network model is locally trained and updated to obtain an updated local model, wherein the weight of the updated local model obtained after training and updating of the local model of the kth computing node is w k (r+1);

[0058] A pruning module, wherein the pruning module prunes the updated local model of each computing node using an adaptive model pruning to obtain a pruned local model;

[0059] An aggregation module, wherein the aggregation module uploads the weights of the local models of each computing node after pruning to the server via an uplink for aggregation to obtain a global model, and the weight of the global model obtained through the aggregation is represented as w(r+1);

[0060] In which, based on the federated learning, the radio modulated signal of the device sequentially passes through the local model training module, the pruning module and the aggregation module to perform multiple rounds of loop functions. After the rth round, the signal is repeatedly input into the local model training module, the pruning module and the aggregation module to proceed to the next round, that is, the r+1th round, until the performance change of the local model and the global model tends to zero, and the global model features and local model features of the feature vector of the serially connected radio modulated signal are used for signal modulation identification to obtain a modulation classification result, wherein the global model features at least include global model weights, and the local model features at least include local model weights.

[0061] In aspect 9, a computing node is provided, characterized in that the method described is run thereon.

[0062] In aspect 10, a computer-readable storage medium is provided, wherein a computer program is stored on the computer-readable storage medium, and the computer program implements the method described when executed by a processor.

[0063] The present invention receives the current state of the global model through a computing node, and performs local updates on this basis, including training and model pruning. Then the computing node sends the updated model parameters (such as weights) back to the server, and the server aggregates the updates of all computing nodes to update the state of the global model. Through multiple rounds of repeated federated learning, until the performance change of the local model and the global model approaches zero, signal modulation recognition is performed according to the global model feature vector and the local model feature vector of the radio modulation signal, and the modulation classification result is obtained, so as to achieve heterogeneous data, solve the challenge of data presenting non-independent and identically distributed, improve the training effect and computing efficiency of the model, and make the information link transmission more efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 Flowchart of the federated learning modulation recognition method for heterogeneous data.

[0065] Figure 2 It is a federated learning training process for heterogeneous data.

[0066] Figure 3 Build a module structure diagram for signal preprocessing and network model.

[0067] Figure 4 This is the structural diagram of the DRSN module.

[0068] Figure 5 This is the structural diagram of the GRU unit.

[0069] Figure 6 This is a comparison chart of the recognition performance of different federated learning frameworks on the test set. DETAILED DESCRIPTION

[0070] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0071] To facilitate understanding of the embodiments of the present invention, the technical terms are first explained. In the communication network of the federated learning application, the edge client device is also called a computing node, and the client runs on the computing node.

[0072] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0073] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0074] The present invention includes embodiments of the first aspect:

[0075] Figure 1 A flowchart of a federated learning modulation recognition method for heterogeneous data is shown. Figure 2 The federated learning training process for heterogeneous data is shown and applied in the federated learning modulation recognition method for heterogeneous data.

[0076] like Figure 1 As shown, a federated learning modulation recognition method for heterogeneous data includes the following steps:

[0077] S1: Receive radio modulated signals, perform signal acquisition and preprocessing on the received signals, and input the preprocessed feature vectors into the federated learning network;

[0078] S2: Federated learning needs to go through multiple rounds of cyclic operations. Assume that the current round is the rth round (r = 1, 2, 3...), and the termination round can also be pre-set according to the actual computing power;

[0079] like Figure 2 As shown, the federated learning network includes k computing nodes, where k=1…K, K represents the number of computing nodes in the federated learning network, and the rth round of federated learning begins. The kth computing node receives the weight w(r) of the global model in the rth round of communication, then the weight of the local model of the kth computing node is w k (r), after completing r rounds of local training, the local model of the k-th computing node obtains the weight w of the updated local model of the k-th computing node k (r+1);

[0080] First, the updated weight of the kth computing node is w k The local model of (r+1) needs to calculate the similarity between the weights of each layer and the weights of the corresponding layer in the global model with weight w(r) It can be expressed as:

[0081]

[0082] in, represents the updated weight of the lth layer of node k after local training in the rth round of training, w l (r) represents the weight of the received global model layer l, where l is a positive integer from 1 to L, L represents the total number of layers in the network, and k and r are positive integers;

[0083] Next, calculate the pruning rate of the lth layer of node k Defined as:

[0084]

[0085] Among them, γ is the preset hyperparameter,

[0086] Finally, according to the obtained pruning rate Calculate the weight of the lth layer Threshold The mask matrix m is derived k (r+1), when the weight value of the lth layer When it is lower than the threshold, the corresponding value of the mask matrix needs to be set to 0, otherwise it is set to 1, which can be expressed as:

[0087]

[0088] The final mask matrix obtained is m k (r+1), when the local model is pruned, it is represented as Where ⊙ represents Hadamard convolution. After the pruning of the computing node k in the rth round of training, the weight of the local model is updated to w k (r+1);

[0089] S3: Upload the weights of the local model of each computing node after pruning to the server via the uplink, such as Figure 2 As described above, the server receives the weight w of the local model k (r+1) is averaged and aggregated to obtain a global model. The weight of the global model obtained by the aggregation is expressed as w(r+1). The process can be defined as:

[0090]

[0091] After the rth round of federated learning is completed, steps S1, S2, and S3 are repeated to proceed to the next round until the performance changes of the local model and the global model tend to zero, and the global model feature vector and the local model feature vector of the feature vector of the serially connected radio modulation signal are used for signal modulation identification to obtain a modulation classification result; the global model feature vector at least includes a global model weight, and the local model feature vector at least includes a local model weight.

[0092] More specifically, step S1 of the federated learning modulation recognition method for heterogeneous data specifically includes:

[0093] S11: receiving radio modulation signals for signal acquisition and preprocessing;

[0094] The I / Q signal and P value signal of the radio modulation signal are obtained. Signal acquisition and preprocessing are aimed at generating input data for the neural network model to facilitate the extraction of feature vectors in subsequent steps. Since radio signal modulation is the process of converting various baseband signals into modulated signals suitable for channel transmission. Considering the main impairments of complex electromagnetic environment, such as center frequency offset, additive white Gaussian noise, fading, etc., the baseband signal r(t) is expressed as:

[0095] r(t)=h(t)*s(t){e j(2πΔf+θ)}+w(t)#(5)

[0096] Where h(t) represents the channel impulse response, which describes the influence of time and frequency selective fading on the signal during transmission, * represents the convolution operation, s(t) represents the original transmitted signal, {e j(2πΔf+θ)} represents the modulation operation, where j is an imaginary unit, Δf and θ represent the frequency offset and phase offset, respectively, which are used to represent the frequency modulation and phase modulation of the signal, respectively, and w(t) represents additive noise or interference, which is a random signal component that may be introduced during the signal transmission process.

[0097] The baseband signal data obtained by equation (1) is stored and processed in the form of I / Q, where I represents the in-phase component and Q represents the orthogonal component. The I / Q signal is obtained by converting r(t), which can be defined as:

[0098] I=Real(r(t))#(6)

[0099] Q=Imag(r(t))#(7

[0100] Where Real(·) and Imag(·) are the real and imaginary parts of r(t). By further transforming the I / Q signal, we can also obtain the P value signal, that is, the instantaneous phase of the signal, which can be expressed as:

[0101]

[0102] S12: Build a federated learning network model. The entire construction process goes through four parts: one-dimensional convolution, deep residual shrinkage network DSRN, GRU unit, and fully connected layer (Full-Connected Layer, FC).

[0103] The automatic modulation recognition process is essentially a multi-classification recognition task. The constructed federated learning network model can extract and fuse the features of the I / Q data and P value signal phase information of the collected signal. Figure 2 This is the overall structure diagram of the federated learning network model, which includes four parts: one-dimensional convolution, deep residual shrinkage network DSRN, GRU unit, and full-connected layer (FC).

[0104] Wherein, step S12 further includes the following steps:

[0105] S121: The process of executing the one-dimensional convolution part is as follows:

[0106] First, if Figure 3 As shown, the one-dimensional convolution is represented as Conv1d, (2, 32) and Conv1d, (1, 32), which means that the I / Q data and P value with data sizes of 2×1024 and 1×1024 respectively pass through a layer of one-dimensional convolution Conv1d, (2, 32) and Conv1d, (1, 32) with 2 and 1 input channels, 32 output channels, 3 convolution kernel size k, and 2 moving steps s. This process can extract the coupling characteristics of I / Q data and the phase information of P value, and obtain two feature vectors of size 32×512. In order to ensure that the network can use the above two feature information at the same time, it is necessary to splice and integrate the two feature vectors into a single feature vector, and the final output feature size is a feature vector of 32×1024.

[0107] S122: The process of the deep residual shrinkage network DRSN part is as follows:

[0108] The deep residual shrinkage network module (DRSN block) improves the efficiency and performance of the model by combining the deep structure and shrinkage strategy of ResNet, which can effectively alleviate the gradient vanishing problem in deep networks and enable the network to learn deeper feature representations. Therefore, a deeper representation of features can be obtained through stacked DRSN modules, so the feature vector obtained in the previous part is continued to pass through 5 stacked DRSN modules to make the output feature size 32×32, which can greatly improve the training efficiency.

[0109] The structure of the DRSN module can be expressed as Figure 4 As shown in the figure, C×W×1 represents the input feature dimension, C represents the number of channels, W represents the feature width, and 1 represents the feature height. The first DRSN module receives the output from the one-dimensional convolution, and its input feature dimensions are C=32 and W=1024. The DRSN module structure adds a soft threshold function to the original ResNet. After the input passes through two layers of one-dimensional convolution layers (BN, ReLU, Conv1d), it is soft-thresholded to obtain an output. The output is directly added to the original input through a one-dimensional average pooling layer (AvgPool1d) to obtain the output of the DRSN module structure. Among them, the soft threshold method is a key step in the signal denoising process. It can set the features with absolute values ​​below a certain threshold to zero and adjust other features toward zero, which can effectively eliminate the influence of noise-related features.

[0110] The input-output relationship of soft thresholding satisfies:

[0111]

[0112] Among them, x and y represent the input and output feature vectors respectively, and τ represents the threshold, which is essentially set under the attention mechanism. It can be observed that the derivative of the input and output is 0 or 1, which can effectively prevent the gradient from disappearing and exploding. This module can generate the threshold corresponding to each channel feature map. The specific process is to first input the feature vector x to take the absolute value (Absolute), add the GAP block operation to obtain a one-dimensional vector, and then pass through two layers of fully connected layers FC to obtain the channel weight vector z, where the number of neurons in the fully connected layer is equal to the number of channels of the feature vector. Finally, the x vector is passed through the Sigmoid function to scale z to the range of (0,1). The specific process is as follows:

[0113]

[0114] Among them, z c Represents the weight of the cth channel, there are C channels in total, α c Represents the obtained scaling value. Finally, take the average of the eigenvalues ​​of the cth channel of the eigenvector x and multiply it by the corresponding α c, we can get the threshold τ of the channel c , to be applied to the soft threshold function. The specific process can be expressed as:

[0115]

[0116] Among them, i and j represent the feature width and height ordinal of the feature vector respectively.

[0117] S123: The process performed by the GRU unit is as follows:

[0118] GRU uses two gates to control the flow of information, namely the reset gate and the update gate. It has a relatively simple structure and a smaller number of parameters, making it more suitable for deployment in resource-constrained computing nodes, such as mobile devices or embedded devices. In order to capture the contextual relationship of the feature vector, the output from the DRSN module needs to pass through a GRU unit with an input size of 32 and a hidden layer number of 128. Its specific structure is as follows Figure 5 shown.

[0119] Among them, x t represents the input information of the GRU unit at the current time t, that is, the output of DRSN, h t-1 Indicates the hidden state of the previous moment, including the previous information, represents the candidate hidden state, h t Represents the hidden state passed to the next moment, r t and z t Represent the reset gate and update gate states, σ and Tanh represent sigmoid and tanh activation functions, respectively. First, the GRU is constructed by h t-1 and the current input x t Get the status of two doors:

[0120] z t =σ(W z x t +U z h t-1 )#(12)

[0121] r t =σ(W r x t +U r h t-1) #(13)

[0122] Where W and U represent the update gate and reset gate pair x respectively. t and h t-1 The weight matrix W includes W z , W r ,U includes U z , U r After that, xt 、h t-1 And the reset gate state r t Linear transformation and Hadamard convolution can obtain candidate hidden states

[0123]

[0124] Hidden state h t By updating the gate state z t and candidate hidden states Combined to get:

[0125]

[0126] Thus, the context relationship of the input feature vector is obtained.

[0127] S124: The process of the fully connected layer (FC) is as follows:

[0128] The feature vector output by the GRU unit is input into the fully connected layer FC, and then the probability of the feature vector belonging to each category is obtained through the Softmax activation function.

[0129] The above three steps S1, S2, and S3 are defined as a round of communication. In each round of communication, the computing node needs to first receive the current state of the global model, and then perform local updates based on this, including training and model pruning. Subsequently, the computing node sends the updated model parameters (such as weights) or gradients back to the server, and the server aggregates the updates of all computing nodes to update the state of the global model. This process is repeated to the next round until the performance change of the model approaches zero, no longer improves significantly, or the preset number of communication rounds is reached.

[0130] The global model features (including but not limited to weights) and local model features (including but not limited to weights) of the serially connected IQ samples are subjected to signal modulation identification to obtain a signal modulation identification result.

[0131] The present invention also includes an embodiment of the second aspect:

[0132] A federated learning modulation recognition device for heterogeneous data, the device comprising:

[0133] An acquisition and preprocessing module, which receives a radio modulation signal, performs signal acquisition and preprocessing on the received signal, and inputs the preprocessed feature vector into the federated learning network;

[0134] A local model training module, wherein the local model training module performs training and updating of the local model, wherein the federated learning network includes k computing nodes, wherein k=1…K, K represents the number of computing nodes in the federated learning network, and the rth round of federated learning is started. The kth computing node receives the weight w(r) of the global model in the rth round of communication, and the weight of the local model of the kth computing node is w k (r), after completing r rounds of local training, the local model of the k-th computing node obtains the weight w of the updated local model of the k-th computing node k (r+1);

[0135] First, the updated weight of the kth computing node is w k The local model of (r+1) needs to calculate the similarity between the weights of each layer and the weights of the corresponding layer in the global model with weight w(r) It can be expressed as:

[0136]

[0137] in, represents the updated weight of the lth layer of node k after local training in the rth round of training, w l (r) represents the weight of the received global model layer l, where l is a positive integer from 1 to L, L represents the total number of layers in the network, and k and r are positive integers;

[0138] Next, calculate the pruning rate of the lth layer of node k Defined as:

[0139]

[0140] Among them, γ is the preset hyperparameter,

[0141] Finally, according to the obtained pruning rate Calculate the weight of the lth layer Threshold The mask matrix m is derived k (r+1), when the weight value of the lth layer When it is lower than the threshold, the corresponding value of the mask matrix needs to be set to 0, otherwise it is set to 1, which can be expressed as:

[0142]

[0143] The final mask matrix obtained is m k (r+1), when the local model is pruned, it is represented as Where ⊙ represents Hadamard convolution. After the pruning of the computing node k in the rth round of training, the weight of the local model is updated to w k (r+1);

[0144] The aggregation module uploads the weights of the local models of each computing node after pruning to the server via an uplink. The server calculates the weights of the local models received by the server. k (r+1) is averaged and aggregated to obtain a global model. The weight of the global model obtained by the aggregation is expressed as w(r+1). The process can be defined as:

[0145]

[0146] After the device performs the rth round of federated learning, the acquisition and preprocessing module, the local model training module and the aggregation module in the device repeat the next round of federated learning until the performance change of the local model and the global model tends to zero, and the global model feature vector and the local model feature vector of the feature vector of the serially connected radio modulation signal are used for signal modulation identification to obtain a modulation classification result; the global model feature vector includes at least a global model weight, and the local model feature vector includes at least a local model weight.

[0147] Furthermore, the local model training module further includes an acquisition and preprocessing submodule and a federated learning network model building submodule;

[0148] Among them, the acquisition and preprocessing submodule is responsible for receiving radio modulation signals for signal acquisition and preprocessing;

[0149] The I / Q signal and P value signal of the radio modulation signal are obtained. Signal acquisition and preprocessing are aimed at generating input data for the neural network model to facilitate the extraction of feature vectors in subsequent steps. Since radio signal modulation is the process of converting various baseband signals into modulated signals suitable for channel transmission. Considering the main impairments of complex electromagnetic environment, such as center frequency offset, additive white Gaussian noise, fading, etc., the baseband signal r(t) is expressed as:

[0150] r(t)=h(t)*s(t){e j(2πΔf+θ)}+w(t)#(5)

[0151] Where h(t) represents the channel impulse response, which describes the influence of time and frequency selective fading on the signal during transmission, * represents the convolution operation, s(t) represents the original transmitted signal, {e j(2πΔf+θ)} represents the modulation operation, where j is an imaginary unit, Δf and θ represent the frequency offset and phase offset, respectively, which are used to represent the frequency modulation and phase modulation of the signal, respectively, and w(t) represents additive noise or interference, which is a random signal component that may be introduced during the signal transmission process.

[0152] The baseband signal data obtained by equation (1) is stored and processed in the form of I / Q, where I represents the in-phase component and Q represents the orthogonal component. The I / Q signal is obtained by converting r(t), which can be defined as:

[0153] I=Real(r(t))#(6)

[0154] Q=Imag(r(t))#(7

[0155] Where Real(·) and Imag(·) are the real and imaginary parts of r(t). By further transforming the I / Q signal, we can also obtain the P value signal, that is, the instantaneous phase of the signal, which can be expressed as:

[0156]

[0157] The federated learning network model construction submodule is responsible for constructing the federated learning network model. The entire construction process goes through four parts: one-dimensional convolution, deep residual shrinkage network DSRN, GRU unit, and fully connected layer (Full-Connected Layer, FC).

[0158] Furthermore, the federated learning network model construction submodule further includes the following units:

[0159] A layer of convolution units, the layer of convolution units is used to extract the data features of the I / Q signal and the phase information of the P value, through a layer of one-dimensional convolution Conv1d, (2, 32) and Conv1d, (1, 32) with input channel numbers of 2 and 1, output channel numbers of 32, convolution kernel sizes k of 3, and moving step sizes s of 2, to obtain two feature vectors with data sizes of 2×1024 and 1×1024 respectively, the two feature vectors are spliced ​​and integrated into a single feature vector, and the final output feature vector has a feature size of 32×1024;

[0160] The DRSN module unit continues to pass the feature vector obtained by the convolutional unit of the first layer through 5 stacked DRSN modules, so that the output feature size is a feature vector of 32×32;

[0161] GRU unit, the output from the DRSN module unit needs to pass through a GRU unit with an input size of 32 and a hidden layer number of 128, so that the GRU unit can capture the contextual relationship of the feature vector;

[0162] The fully connected layer FC unit maps the feature vector output by the GRU unit to the final output through a fully connected layer FC, the output size is 1×24, and the probability that the input feature vector belongs to each category is obtained through the Softmax activation function, and the category of the radio modulation signal is determined according to the probability.

[0163] The invention also includes a computing node on which the method described above is executed. The invention also includes a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the method described above is implemented.

[0164] Technical Effects of the Invention

[0165] Following the above technical solution, a fully trained global model is extracted, and the fully trained global model is applied to radio signal modulation recognition. In order to evaluate the technical effect of the technical solution, the classification recognition effect and the ability to reduce computing overhead of the present invention are considered, and key indicators such as recognition accuracy and communication cost are selected for performance evaluation.

[0166] (1) Recognition accuracy

[0167] In this modulation recognition task, the recognition accuracy refers to the proportion of samples correctly predicted by the model to the total number of samples. The calculation formula is:

[0168]

[0169] Among them, TP refers to the positive tuple correctly classified by the classifier; TN refers to the negative tuple correctly classified by the classifier; FP refers to the negative tuple incorrectly marked as a positive tuple by the classifier; FN refers to the positive tuple incorrectly marked as a negative tuple by the classifier.

[0170] (2) Communication cost

[0171] In order to measure the cost and resources required to send and receive information during the entire training process, the calculation formula of the communication cost C defined in the present invention is:

[0172]

[0173] Among them, E represents the number of computing nodes, R represents the total number of training rounds, also known as the number of communication rounds, P(r) represents the pruning rate of the r-th round of pre-pruning training, and P(r+1) represents the pruning rate of the model after the r-th round of training.

[0174] In order to obtain signals of various modulation types, we use multiple general radio peripherals B210 to modulate and transmit signals, and use another B120 to receive signals. In the process of signal simulation generation, the present invention takes into account the influence of signal frequency deviation, phase shift and multipath effect, and finally generates five modulation type signals such as OOK, 4ASK, BPSK, OQPSK and GMSK. The signal-to-noise ratio range is -20 to 30dB. At intervals of 2dB, 1,600 signal samples are generated under each signal-to-noise ratio condition, totaling 208,000 signal samples. It is divided into training set and test set in a ratio of 8:2, that is, 166,400 training samples and 41,600 test samples. Each type of signal in the training set is divided into 4 parts, a total of 20 parts, and the 20 parts are randomly assigned to 10 computing nodes, that is, each signal node has only two types of signals.

[0175] In order to verify the recognition performance of the proposed federated learning framework under heterogeneous data conditions, the proposed framework is compared with the commonly used FedAvg and FedProx frameworks, and the recognition performance of different federated learning frameworks on the test set is obtained as follows: Figure 6 As shown in the figure, ours represents the heterogeneous federated learning framework of the present invention, FedAvg represents the common weighted average federated learning framework, and FedProx represents the common improved federated learning framework. It can be observed that due to the application of the adaptive model pruning method, the local models of different computing nodes can be adjusted according to their respective data distributions, which is more suitable for the scenario of heterogeneous data and has a higher recognition accuracy.

[0176] In order to observe the computational efficiency of different federated learning frameworks, their communication costs were obtained and compared under the same training conditions, where the number of computing nodes N was equal to 10. Finally, the number of communication rounds of the method of the present invention was 20, the number of communication rounds of FedAvg was 40, and the number of communication rounds of FedProx was 36. The pruning rate of the present invention is greater than 0. Since FedAvg and FedProx do not use the model pruning method for pruning, their pruning rates can be regarded as 0, and their communication costs are obtained as follows:

[0177]

[0178] C FedAvg =10×40×2=800#(19)

[0179] C FedProx =10×36×2=720#(20)

[0180] Obviously C ours <C FedProx <C FedAvg , so the proposed method has the smallest communication cost and the highest computational efficiency.

[0181] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Although the present invention has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the technical solutions described in the above embodiments can still be modified, or some or all of the technical features therein can be replaced by equivalents, and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A federated learning modulation recognition method for heterogeneous data, characterized in that: The following steps are involved: S1: Receive radio modulated signals, perform signal acquisition and preprocessing on the received signals, and input the preprocessed feature vectors into the federated learning network; S2: The federated learning network includes k computing nodes, where k=1…K, K represents the number of computing nodes in the federated learning network. The rth round of federated learning begins. The kth computing node receives the weight w(r) of the global model in the rth round of communication. Then the weight of the local model of the kth computing node is w k (r), after completing r rounds of local training, the local model of the k-th computing node obtains the weight w of the updated local model of the k-th computing node k (r+1); First, the updated weight of the kth computing node is w k The (k+1) local model needs to calculate the similarity between the weights of each layer and the weights of the corresponding layer in the global model with weight w(r) It can be expressed as: in, represents the updated weight of the lth layer of node k after local training in the rth round of training, w l (r) represents the weight of the received global model layer l, where l is a positive integer from 1 to L, L represents the total number of layers in the network, and k and r are positive integers; Next, calculate the pruning rate of the lth layer of node k Defined as: Among them, γ is the preset hyperparameter, Finally, according to the obtained pruning rate Calculate the weight of the lth layer Threshold The mask matrix m is derived k (r+1), when the weight value of the lth layer When it is lower than the threshold, the corresponding value of the mask matrix needs to be set to 0, otherwise it is set to 1, which can be expressed as: The final mask matrix obtained is m k (r+1), when the local model is pruned, it is represented as Where ⊙ represents Hadamard convolution. After the pruning of the computing node k in the rth round of training, the weight of the local model is updated to w k (r+1); S3: The weights of the local models of each computing node after pruning are uploaded to the server via the uplink. The server calculates the weights w of the received local models. k (r+1) is aggregated to obtain a global model. The weight of the global model obtained by the aggregation is expressed as w(r+1), which is specifically defined as: After the rth round of federated learning is completed, steps S1, S2, and S3 are repeated to proceed to the next round until the performance changes of the local model and the global model tend to zero, and the global model feature vector and the local model feature vector of the feature vector of the serially connected radio modulation signal are used to perform radio modulation signal identification to obtain a modulation classification result; the global model feature vector at least includes a global model weight, and the local model feature vector at least includes a local model weight.

2. The method for federated learning modulation recognition for heterogeneous data according to claim 1, characterized in that: The step S1 further specifically includes the following steps: S11: receiving radio modulation signals for signal acquisition and preprocessing; S12: Build a federated learning network model. The entire construction process goes through four parts: one-dimensional convolution, deep residual shrinkage network DSRN, GRU unit, and fully connected layer (Full-Connected Layer, FC).

3. The method for federated learning modulation recognition for heterogeneous data according to claim 2, characterized in that: Wherein the step S12 further comprises: Through a layer of one-dimensional convolution Conv1d,(2,32) and Conv1d,(1,32) with input channel numbers of 2 and 1, output channel numbers of 32, convolution kernel sizes k of 3, and moving step sizes s of 2, the data features of the I / Q signal and the phase information of the P-value signal are extracted, and two feature vectors with data sizes of 2×1024 and 1×1024 are obtained, wherein the I / Q signal is a signal stored and processed in the form of I / Q of the baseband signal of the radio modulation signal obtained from signal acquisition and preprocessing, wherein I represents the in-phase component, Q represents the orthogonal component, and the P-value signal represents the instantaneous phase of the baseband signal of the radio modulation signal; the two feature vectors are spliced ​​and integrated into a single feature vector, and the feature vector with a data size of 32×1024 is output; Continue to pass the feature vector with a data size of 32×1024 through 5 stacked DRSN modules to output a feature vector with a data size of 32×32; Next, the output from the DRSN module passes through the GRU unit so that the GRU unit captures the contextual relationship of the feature vector; Finally, the feature vector output by the GRU unit is input into the fully connected layer FC, and then the probability of the feature vector belonging to each category is obtained through the Softmax activation function.

4. The method for federated learning modulation recognition for heterogeneous data according to claim 3, characterized in that: The GRU unit captures the contextual relationship of the feature vector and further include: Among them, use x t represents the input information of the GRU unit at the current time t, that is, the output of the DRSN module, h t-1 Indicates the hidden state of the previous moment, including the previous information, represents the candidate hidden state, h t Represents the hidden state passed to the next moment, including the subsequent information of the feature vector, r t and z t Represent the reset gate and update gate states respectively, σ and Tanh represent sigmoid and tanh activation functions respectively; the GRU unit is activated by h t-1 and the current input x t Get the status of two doors: z t =σ(W z x t +U z h t-1) #(5) r t =σ(W r x t +U r h t-1 )#(6) Where W and U represent the update gate and reset gate for x respectively. t and h t-1 The weight matrix W includes W z , W r , U includes U z , U r , then, for x t 、h t-1 And the reset gate state r t Linear transformation and Hadamard convolution can obtain candidate hidden states Finally, the hidden state h t By updating the gate state z t and candidate hidden states Combined to get: Thus, the context relationship of the input feature vector is obtained.

5. A federated learning modulation recognition device for heterogeneous data, characterized in that: The device comprises the following: A signal preprocessing and network model building module, wherein the signal preprocessing and network model building module receives a radio modulation signal, performs signal acquisition and preprocessing on the received signal, inputs the preprocessed feature vector into the federated learning network, and builds a network model; A local model training module, wherein the local model training module performs training and updating of the local model, wherein the federated learning network includes k computing nodes, wherein k=1…K, K represents the number of computing nodes in the federated learning network, and the rth round of federated learning is started. The kth computing node receives the weight w(r) of the global model in the rth round of communication, and the weight of the local model of the kth computing node is w k (r), after completing r rounds of local training, the local model of the k-th computing node obtains the weight w of the updated local model of the k-th computing node k (r+1); First, the updated weight of the kth computing node is w k The (k+1) local model needs to calculate the similarity between the weights of each layer and the weights of the corresponding layer in the global model with weight w(r) It can be expressed as: in, represents the updated weight of the lth layer of node k after local training in the rth round of training, w l (r) represents the weight of the received global model layer l, where l is a positive integer from 1 to L, L represents the total number of layers in the network, and k and r are positive integers; Next, calculate the pruning rate of the lth layer of node k Defined as: Among them, γ is the preset hyperparameter, Finally, according to the obtained pruning rate Calculate the weight of the lth layer Threshold The mask matrix m is derived k (r+1), when the weight value of the lth layer When it is lower than the threshold, the corresponding value of the mask matrix needs to be set to 0, otherwise it is set to 1, which can be expressed as: The final mask matrix obtained is m k (r+1), when the local model is pruned, it is represented as Where ⊙ represents Hadamard convolution. After the pruning of the computing node k in the rth round of training, the weight of the local model is updated to w k (r+1); The aggregation module uploads the weights of the local models of each computing node after pruning to the server via an uplink. The server calculates the weights of the local models received by the server. k (r+1) is averaged and aggregated to obtain a global model. The weight of the global model obtained by the aggregation is expressed as w(r+1), which is specifically defined as: After the device performs the rth round of federated learning, the acquisition and preprocessing module, the local model training module and the aggregation module in the device repeat the next round of federated learning until the performance change of the local model and the global model tends to zero, and the global model feature vector and the local model feature vector of the feature vector of the serially connected radio modulation signal are used to perform radio modulation signal identification to obtain a modulation classification result; the global model feature vector at least includes a global model weight, and the local model feature vector at least includes a local model weight.

6. The federated learning modulation recognition device for heterogeneous data according to claim 5, characterized in that: The signal preprocessing and network model building module further includes an acquisition and preprocessing submodule and a federated learning network model building submodule; Among them, the acquisition and preprocessing submodule is responsible for receiving radio modulation signals for signal acquisition and preprocessing; The federated learning network model construction submodule is responsible for constructing the federated learning network model. The entire construction process goes through four parts: one-dimensional convolution, deep residual shrinkage network DSRN, GRU unit, and fully connected layer (Full-Connected Layer, FC).

7. The federated learning modulation recognition device for heterogeneous data according to claim 6, characterized in that: The federated learning network model construction submodule also includes the following units: A layer of convolution units, the layer of convolution units is used to extract the data features of the I / Q signal and the phase information of the P value, through a layer of one-dimensional convolution Conv1d, (2, 32) and Conv1d, (1, 32) with input channel numbers of 2 and 1, output channel numbers of 32, convolution kernel sizes k of 3, and moving step sizes s of 2, to obtain two feature vectors with data sizes of 2×1024 and 1×1024 respectively, the two feature vectors are spliced ​​and integrated into a single feature vector, and the final output feature vector has a feature size of 32×1024; The DRSN module unit continues to pass the feature vector obtained by the convolutional unit of the first layer through 5 stacked DRSN modules, so that the output feature size is a feature vector of 32×32; GRU unit, the output from the DRSN module unit needs to pass through a GRU unit with an input size of 32 and a hidden layer number of 128, so that the GRU unit can capture the contextual relationship of the feature vector; The fully connected layer FC unit maps the feature vector output by the GRU unit to the final output through a fully connected layer FC, the output size is 1×24, and the probability that the input feature vector belongs to each category is obtained through the Softmax activation function, and the category of the radio modulation signal is determined according to the probability.

8. The federated learning modulation recognition device for heterogeneous data according to claim 7, characterized in that: The GRU unit captures the contextual relationship of the feature vector and further include: Among them, use x t represents the input information of the GRU unit at the current time t, that is, the output of the DRSN module, h t-1 Indicates the hidden state of the previous moment, including the previous information, represents the candidate hidden state, h t Represents the hidden state passed to the next moment, including the subsequent information of the feature vector, r t and z t Represent the reset gate and update gate states respectively, σ and Tanh represent sigmoid and tanh activation functions respectively; the GRU unit is activated by h t-1 and the current input x t Get the status of two doors: z t =σ(W z x t +U z h t-1 )#(5) r t =σ(W r x t +U r h t-1 )#(6) Where W and U represent the update gate and reset gate for x respectively. t and h t-1 The weight matrix W includes W z , W r , U includes U z , U r , then, for x t 、h t-1 And the reset gate state r t Linear transformation and Hadamard convolution can obtain candidate hidden states Finally, the hidden state h t By updating the gate state z t and candidate hidden states Combined to get: Thus, the context relationship of the input feature vector is obtained.

9. A computing node, characterized in that: The method according to any one of claims 1 to 4 is carried out thereon.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images