A modulation recognition optimization method for communication signals based on signal parameter estimation

Abstract

The present invention discloses a modulation recognition optimization method for communication signals based on signal parameter estimation. First, a training data set is constructed, and then a multi-task deep convolutional neural network for signal parameter recognition is constructed and trained. Then, a deep convolutional neural network for signal modulation recognition is constructed and trained to obtain a pre-trained model. Then, a fine-tune is performed within a small range of parameter data sets. Finally, a test sample set is used to obtain the modulation recognition accuracy after parameter estimation optimization. The method of the present invention has a faster reasoning speed after being deployed using RepVGG as the backbone structure. The deep convolutional neural network is used to identify signal parameters, and the results are fully utilized in further training and processing. This enables the modulation recognition task to make full use of the parameter information of the signal, rather than just relying on the waveform information of the signal for feature extraction. In addition, the use of fine-tune can maintain a similar or even better signal modulation recognition effect than retraining in the task of downlink modulation recognition with less training time.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications, and in particular relates to a modulation identification optimization method for communication signals based on signal parameter estimation. Background Art

[0002] Signal modulation identification (AMI) is a method for identifying the modulation type of radio signals. AMI determines the modulation type by analyzing the waveform and spectral characteristics of the received signal. AMI is a key method for distinguishing communication signal types and plays an important role in civilian communications, such as spectrum monitoring and interference signal detection.

[0003] Existing modulation classification algorithms are model-driven and are generally categorized into likelihood-based and expert feature extraction-based algorithms. In existing CDMA communication systems, wireless links have control channels specifically used to transmit control information. These primarily include pilot channels and synchronization channels. The pilot channel is a pilot signal transmitted continuously or on demand by the base station. Its primary function is to provide a coherent carrier for demodulation by mobile stations and serve as a measurement signal for handovers. The synchronization channel is a synchronization signal continuously transmitted by the base station, providing synchronization information to the receiving mobile station.

[0004] However, in non-cooperative communications, the receiver cannot obtain the precise signal parameters of the transmitter through pilot signals. It must rely on blind signal processing techniques to determine parameters such as the carrier frequency and modulation scheme for subsequent demodulation. In this process, pre-estimating the signal's carrier frequency, symbol rate, and degree of distortion (carrier frequency deviation) to narrow the range is crucial for subsequent modulation scheme identification.

[0005] CN116068520A discloses a Transformer-based cognitive radar joint modulation identification and parameter estimation method. This method can simultaneously automatically identify the modulation type and estimate the modulation parameters of variable structures for radar operating modes defined by different modulation types and modulation parameter combinations of various control parameters. Through specialized label sequence design, radar signal modulation type identification and signal parameter estimation are treated as a single NLP task, with sequential output testing the radar signal category and other parameters. However, this method does not further verify that multi-task parameter sharing training tasks combining modulation category and parameter estimation indicators can improve the accuracy of modulation category. Summary of the Invention

[0006] To solve the above technical problems, the present invention proposes a modulation recognition optimization method for communication signals based on signal parameter estimation, which is used to solve the problem that the parameter estimation index is not highly correlated with the accuracy of signal modulation category recognition in complex channel environments. By performing parameter estimation on the data set and obtaining a pre-trained model in the modulation recognition task, fine-tuning can achieve results similar to those of retraining.

[0007] The technical solution adopted by the present invention is: a modulation recognition optimization method for communication signals based on signal parameter estimation, the specific steps are as follows:

[0008] S1. Build a training dataset;

[0009] S2. Build and train a multi-task deep convolutional neural network for signal parameter recognition;

[0010] S3. Build and train a deep convolutional neural network for signal modulation recognition to obtain a pre-trained model, and then fine-tune it within a small parameter dataset.

[0011] S4. Use the test sample set to obtain the modulation recognition accuracy after parameter estimation optimization.

[0012] Furthermore, the step S1 is specifically as follows:

[0013] S11. Use MATLAB simulation software to simulate a general communication link structure and generate a data set for signal parameter identification and modulation identification;

[0014] Two datasets for signal parameter identification and modulation identification were generated in the same way: the first dataset was used for signal parameter estimation and signal modulation identification tasks, while the second dataset was only used to pre-train the deep convolutional neural network model for fine-tuning signal modulation identification. The two datasets differed only in the signal symbol rate range and frequency offset range.

[0015] Among them, the data sets used to generate the pre-training models in the signal parameter estimation task and the signal modulation recognition task are the same, both using the first data set. The signal parameter estimation task uses the signal symbol rate and frequency offset as labels, and the signal modulation recognition task uses the modulation category as the label.

[0016] S12, generates signals of 12 modulation types;

[0017] In step S11 , the data volume and channel conditions of the two data sets are the same, and the only difference is the range of symbol rate and the range of frequency offset.

[0018] Each modulated signal contains 20,000 data items with a signal-to-noise ratio of 0-20dB, for a total of 240,000 data items;

[0019] Of these, 80% were used as training samples, 10% as validation samples, and 10% as testing samples. For the first dataset, the symbol rate varied randomly between 100kHz and 200kHz, showing a uniform distribution, and the frequency offset varied randomly between -10% and 10%, showing a uniform distribution. For the second dataset, the symbol rate varied randomly between 93kHz and 107kHz, showing a uniform distribution, and the frequency offset varied randomly between -2% and 2%, showing a uniform distribution.

[0020] The channel conditions for generating data are set to Gaussian channel and Rayleigh channel; finally, the IQ values ​​of the signal are stored as a data set.

[0021] S13, filtering the two data sets in step S11 through a low-pass filter and normalizing them;

[0022] The data set is filtered through a Chebyshev low-pass filter with a set cutoff frequency, and then normalized to its maximum and minimum values. The normalization formula is as follows:

[0023]

[0024] Among them, Y i represents the i-th signal sample X i The sample value after normalization. Max represents the maximum value of a single sample data, and Min represents the minimum value of a single sample data. Signal labels include: modulation type, symbol rate, and frequency offset.

[0025] Furthermore, the step S2 is specifically as follows:

[0026] S21. Construct a multi-task deep convolutional neural network for signal parameter recognition based on the repvgg backbone structure;

[0027] Signal parameter recognition multi-task deep convolutional neural network, its structure includes: input layer, backbone network layer, feature extraction layer, classification layer; and the network adopts loss function L mixed Measure the error between the model prediction and the true target and optimize the model parameters; specifically as follows:

[0028] (1) Input layer: The input dimension of the network is (64, 2, 10000);

[0029] The first dimension represents the number of data in the data block of the data input, batch_size; the second dimension represents the two IQ channels; and the third dimension represents the signal length of each IQ dimension.

[0030] (2) Backbone network layer: adopts the repvgg block stacking structure and outputs the feature vector through the global average pooling layer GAP;

[0031] The repvgg block mainly consists of two parts:

[0032] Part 1: After the data is input, there are three branches, and then the three branches are superimposed together. The first branch consists of a convolution layer and a batch normalization layer, the second branch consists of a 1×1 convolution and a BN layer, and the third branch directly passes the data through the BN layer to the next layer; after the three branches are superimposed, they pass through the rectified linear unit ReLU; and only the first branch is multiplied by the learnable channel weight matrix M weight The working principle of this weight is expressed as follows:

[0033]

[0034] Among them, λ l,C The subscript C represents the number of channels of the first branch data input, and l represents the current depth of the repvgg block, that is, the network is composed of D repvgg blocks stacked together. The network structure expression of the first part is as follows:

[0035]

[0036] Among them, x l ′ represents the output of the first three branches in the lth repvgg block; The convolutional neural network representing the first branch of the first part of the repvgg block, x l Represents the input of the first part of the repvgg block and serves as the third branch of this part.

[0037] Part 2: Input the superimposed features into a data channel adaptive weight block (SE-Block), that is, multiply each channel of the input data features by the weight value obtained by a channel weight extraction layer;

[0038] Among them, the specific composition of the channel weight extraction layer is: global average pooling layer GAP, 1×1 convolution layer, rectified linear unit ReLU, and 1×1 convolution layer.

[0039] The above two parts are combined and stacked to form a backbone network.

[0040] (3) Feature extraction layer: The feature vector obtained by the backbone network is split into two channels, which are used for the regression tasks of symbol rate fd and signal frequency offset offset respectively;

[0041] The feature extraction layer structure used in the above two tasks is the same, both adopt a three-layer fully connected structure. The first two layers have rectified linear units (ReLUs), and the first layer is set with a dropout layer with a dropout parameter of 0.15. Finally, the feature vectors used for signal parameter regression are output respectively.

[0042] (4) Classification layer: corresponds to two parallel feature extraction networks, and the activation function of the final output layer changes according to the regression parameters;

[0043] Among them, for the symbol rate fd: its normalized label range is [0, 1], so sigmoid is used as the activation function; for the frequency offset offset: its normalized label range is [-1, 1], so Tanh is used as the activation function.

[0044] Through the signal parameter recognition multi-task deep convolutional neural network can simultaneously regress and obtain the predicted values ​​of symbol rate and frequency offset.

[0045] S22. Set the loss function of the signal parameter recognition multi-task deep convolutional neural network to be the sum of two mean square errors multiplied by the corresponding weights, i.e., L mixed The network parameter update algorithm is the error back propagation algorithm. The activation function corresponds to different regression tasks and is set to Sigmoid and Tanh respectively. The Adam optimizer is used for optimization. The initial learning rate lr is set to 0.0002, Adam∈ is set to 1e-8, Adamβ is set to (0.9, 0.999), weight decay is set to 0, and the lr update criterion is warmupcosinedecay.

[0046] Among them, L mixed It consists of two parts, namely the mean square error of the two tasks multiplied by the corresponding weights:

[0047]

[0048] in, represents the mean square error of the fd task, represents the mean square error of the offset task; μ1 and μ2 represent the weights of the mean square error of the fd task and the offset task, which are set to 0.3 and 0.7 during training.

[0049] The main parameters of the learning rate update criterion of Warmupcosine include: the number of warm-up training epochs warmup_epochs, the maximum learning rate base_value of the cosine learning rate curve, the minimum learning rate final_value, the period T_epochs of the cosine learning rate curve, and the decay of the cosine learning rate curve decay, that is, after each training epoch T_epochs, the maximum value of the learning rate curve will be multiplied by decay, and the number of iterations of each epoch n_iter.

[0050] The final learning curve is: at the beginning of the warm-up training, it increases linearly from 0 to the maximum value, then decreases from the maximum value to the minimum value in one cycle of cosine, and then jumps to the maximum value after decay, and so on until the training is completed.

[0051] S23, shuffle the training samples and input them into the network for training;

[0052] The network is trained using two parallel multi-task labels, symbol rate fd and frequency offset offset, according to the loss function and optimization algorithm of step S22.

[0053] S24, testing the signal parameter recognition effect of the trained network;

[0054] The IQ data signal of the test sample set is input into the signal parameter recognition multi-task deep convolutional neural network, and the mean absolute error (MAE) is used to measure the regression performance of the network's symbol rate fd and frequency offset offset to obtain the parameter recognition result.

[0055] Furthermore, the step S3 is specifically as follows:

[0056] S31. Construct a deep convolutional neural network structure for signal modulation recognition based on the repvgg backbone structure;

[0057] Construct a deep convolutional neural network structure for signal modulation recognition based on the repvgg backbone structure, which includes: input layer, backbone network layer, feature extraction layer, classification layer; and the network adopts the loss function binary cross entropy L CrossEntropyLoss Measure the error between the model prediction and the true target and optimize the model parameters; specifically as follows:

[0058] (1) Input layer: The input dimension of the network is (64, 2, 10000);

[0059] The first dimension represents the number of data in the data block of the data input, batch_size; the second dimension is the two IQ channels; and the third dimension is the signal length of each IQ dimension.

[0060] (2) Backbone network layer: the same as the backbone network of the multi-task deep convolutional neural network for signal parameter recognition in step S21;

[0061] (3) Feature extraction layer: The first layer of the feature extraction layer is the global average pooling layer GAP, and then the last dimension of the average pooling is removed. It passes through four consecutive layers of fully connected linear layers, the first three layers of rectified linear units ReLU, and the first three layers are all set with a dropout layer Dropout, with the dropout parameter set to 0.15, and finally outputs the feature vector of the number of categories;

[0062] The feature vectors obtained by the backbone network are used for the regression tasks of symbol rate fd and signal frequency offset offset.

[0063] The feature extraction layer structures used in the above two tasks are the same; both adopt a three-layer fully connected structure, the first two layers have rectified linear units (ReLU), and the first layer is set with a dropout layer (Dropout), the dropout parameter is set to 0.15, and the final output is the feature vector of the number of categories.

[0064] (4) Classification layer: Modulation recognition is a multi-classification task, and Softmax is used as the activation function.

[0065] S32. Set the loss function of the signal modulation recognition deep convolutional neural network to binary cross entropy L CrossEntropyLoss The network parameter update algorithm is the error back propagation algorithm, the activation function is softmax, and the Adam optimizer is used for optimization. The initial learning rate lr is set to 0.0002, Adam∈ is set to 1e-8, Adamβ is set to (0.9, 0.999), weight decay is set to 0, and the lr update criterion is warmupcosinedecay;

[0066] Loss function binary cross entropy L CrossEntropyLoss The expression is as follows:

[0067]

[0068] Where M represents the total number of signal modulation categories, N represents the number of samples in a batch; w m represents the weight of each category, x n,i Indicates the value of the nth data in the i-th label dimension, y n,m Represents the one-hot encoded label vector, which represents the value of the i-th category dimension of the n-th data label. If category x n is category m, then y n,m is 1, and other i not equal to m are 0.

[0069] The learning rate update criterion warmupcosine is the same as that in step S22.

[0070] S33, training on the first data set described in step S11 to obtain a pre-trained model;

[0071] S34, fine-tuning the pre-trained model on the dataset with the parameter range narrowed down according to the result of parameter identification in step S24, i.e., the second dataset in step S11;

[0072] Use the fine-tune strategy of DeiT (Data-Efficient Image Transformer) in the image field:

[0073] The algorithm is fully trained on a 224×224 resolution image dataset and then fine-tuned on a high-resolution 384×384 dataset, achieving results similar to or even better than retraining on 384×384 resolution data.

[0074] Furthermore, the step S4 is specifically as follows:

[0075] The IQ data signal of the test sample set is input into the signal modulation recognition deep convolutional neural network, the position of the maximum result in the softmax output label dimension is used as the output modulation category result, and the accuracy is used to measure the network modulation recognition performance.

[0076] Beneficial effects of the present invention: The method of the present invention first constructs a training data set, then constructs and trains a multi-task deep convolutional neural network for signal parameter recognition, and then constructs and trains a deep convolutional neural network for signal modulation recognition to obtain a pre-trained model, and then fine-tunes within a small range of parameter data sets, and finally uses a test sample set to obtain the modulation recognition accuracy after parameter estimation optimization. The method of the present invention uses a network structure for signal parameter recognition and signal modulation category recognition based on a deep convolutional neural network with a RepVGG backbone structure, which overcomes the shortcomings of the existing method in that the process of manually extracting signal features is cumbersome and requires a large amount of expert prior knowledge to complete signal recognition. Due to the convolutional layer setting of the three branches in the RepVGG block, during the reasoning process, the three branches can be merged into one convolutional layer operation. Therefore, during the reasoning process, compared with the convolutional neural network with a general Inception structure, the method of the present invention has a faster reasoning speed, uses a deep convolutional neural network to recognize signal parameters, and fully utilizes the results for further training and processing, which enables the modulation recognition task to make full use of the signal parameter information rather than relying solely on the signal waveform information for feature extraction. In addition, the use of fine-tune can maintain a signal modulation recognition effect similar to or even better than retraining in the downlink modulation recognition task with less training time. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is a flow chart of a modulation identification optimization method for communication signals based on signal parameter estimation according to the present invention.

[0078] Figure 2 This is a structural diagram of the multi-task deep convolutional neural network for signal parameter recognition used in an embodiment of the present invention.

[0079] Figure 3 This is a structural diagram of the deep convolutional neural network for signal modulation recognition used in an embodiment of the present invention.

[0080] Figure 4 This is a diagram showing the effect of a modulation identification test before and after the parameter identification control parameter range in an embodiment of the present invention.

[0081] Figure 5 This is a comparison chart of the test results of fine-tune and retraining in an embodiment of the present invention and the test results after other network training. DETAILED DESCRIPTION

[0082] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0083] like Figure 1 As shown in FIG, a flow chart of a modulation identification optimization method for communication signals based on signal parameter estimation of the present invention, the specific steps are as follows:

[0084] S1. Build a training dataset;

[0085] S2. Build and train a multi-task deep convolutional neural network for signal parameter recognition;

[0086] S3. Build and train a deep convolutional neural network for signal modulation recognition to obtain a pre-trained model, and then fine-tune it within a small parameter dataset.

[0087] S4. Use the test sample set to obtain the modulation recognition accuracy after parameter estimation optimization.

[0088] In this embodiment, step S1 is specifically as follows:

[0089] S11. Use MATLAB simulation software to simulate a general communication link structure and generate a data set for signal parameter identification and modulation identification;

[0090] Two datasets for signal parameter identification and modulation identification were generated in the same way: the first dataset was used for signal parameter estimation and signal modulation identification tasks, while the second dataset was only used to pre-train the deep convolutional neural network model for fine-tuning signal modulation identification. The two datasets differed only in the signal symbol rate range and frequency offset range.

[0091] Among them, the data sets used to generate the pre-training models in the signal parameter estimation task and the signal modulation recognition task are the same, both using the first data set. The signal parameter estimation task uses the signal symbol rate and frequency offset as labels, and the signal modulation recognition task uses the modulation category as the label.

[0092] S12, generates signals of 12 modulation types;

[0093] In step S11 , the data volume and channel conditions of the two data sets are the same, and the only difference is the range of symbol rate and the range of frequency offset.

[0094] Each modulated signal contains 20,000 data items with a signal-to-noise ratio of 0-20dB, for a total of 240,000 data items;

[0095] Of these, 80% were used as training samples, 10% as validation samples, and 10% as testing samples. For the first dataset, the symbol rate varied randomly between 100kHz and 200kHz, showing a uniform distribution, and the frequency offset varied randomly between -10% and 10%, showing a uniform distribution. For the second dataset, the symbol rate varied randomly between 93kHz and 107kHz, showing a uniform distribution, and the frequency offset varied randomly between -2% and 2%, showing a uniform distribution.

[0096] In order to simulate the noise environment and multipath effect of the actual signal, the channel conditions for generating data are set to Gaussian channel and Rayleigh channel; finally, the IQ two-way values ​​of the signal are stored as a data set.

[0097] S13, filtering the two data sets in step S11 through a low-pass filter and normalizing them;

[0098] The data set is filtered through a Chebyshev low-pass filter with a set cutoff frequency, and then normalized to its maximum and minimum values. The normalization formula is as follows:

[0099]

[0100] Among them, Y i represents the i-th signal sample X i The sample value after normalization. Max represents the maximum value of a single sample data, and Min represents the minimum value of a single sample data. Signal labels include: modulation type, symbol rate, and frequency offset.

[0101] In this embodiment, step S2 is specifically as follows:

[0102] S21. Construct a multi-task deep convolutional neural network for signal parameter recognition based on the repvgg backbone structure;

[0103] Signal parameter recognition multi-task deep convolutional neural network structure such as Figure 2 As shown, its structure includes: input layer, backbone network layer, feature extraction layer, classification layer; and the network adopts loss function L mixed Measure the error between the model prediction and the true target to optimize the model parameters; specifically as follows:

[0104] (1) Input layer: The input dimension of the network is (64, 2, 10000);

[0105] The first dimension represents the number of data in the data block of the data input, batch_size, which is set to 64 in this embodiment. The second dimension represents the two IQ channels. The third dimension represents the signal length of each IQ dimension.

[0106] (2) Backbone network layer: adopts repvgg block stacking structure, Figure 2 Shown is a repvgg block N×, which outputs a feature vector through a global average pooling layer GAP;

[0107] The repvgg block mainly consists of two parts:

[0108] Part 1: After the data is input, there are three branches, and then the three branches are superimposed together. The first branch consists of a convolution layer and a batch normalization layer (BN layer), the second branch consists of a 1×1 convolution and a BN layer, and the third branch is directly fed by the data through the BN layer into the next layer; after the three branches are superimposed, they pass through the rectified linear unit ReLU; and only the first branch is multiplied by the learnable channel weight matrix M weight .exist Figure 2 The weight matrix block is represented in the figure. The weight principle expression is as follows:

[0109]

[0110] Among them, λ l,C The subscript C represents the number of channels of the first branch data input, and l represents the current depth of the repvgg block, that is, the network is composed of D repvgg blocks stacked together. The weight matrix M weight It is a diagonal matrix, and the parameters on the diagonal are all learnable, that is, they are updated as the network model parameters are updated. The network structure expression of the first part is as follows:

[0111]

[0112] Among them, x l ′ represents the output of the first three branches in the lth repvgg block; The convolutional neural network representing the first branch of the first part of the repvgg block, x l It represents the input of the first part of the repvgg block and also serves as the third branch of this part.

[0113] Part 2: Input the superimposed features into a data channel adaptive weight block (SE-Block), that is, multiply each channel of the input data features by the weight value obtained by a channel weight extraction layer;

[0114] The specific composition of the channel weight extraction layer is: global average pooling layer GAP, 1×1 convolution layer, rectified linear unit ReLU, 1×1 convolution layer; the channel weight extraction layer and the channel weight matrix M of the first part weight They have the same function and can update weights as training progresses. The difference is that the weight matrix is ​​updated directly, while the channel weight extraction layer needs to be calculated through a convolutional neural network. Its mechanism is similar to self-attention, and both give greater weights to channels with larger values, that is, focus on channels with prominent data as extracted features.

[0115] Combining and stacking the above two parts will form the backbone network.

[0116] (3) Feature extraction layer: Due to the use of a multi-task structure, the feature vector obtained by the backbone network is split into two channels, which are used for the regression tasks of symbol rate fd and signal frequency offset offset respectively;

[0117] The feature extraction layer structure used in the above two tasks is the same, both adopt a three-layer fully connected structure. The first two layers have rectified linear units (ReLU), and the first layer is set with a dropout layer (Dropout) with a dropout parameter of 0.15. Finally, the feature vectors used for signal parameter regression are output respectively.

[0118] (4) Classification layer: corresponds to two parallel feature extraction networks, and the activation function of the final output layer changes according to the regression parameters;

[0119] Among them, for the symbol rate fd: its normalized label range is [0, 1], so sigmoid is used as the activation function; for the frequency offset offset: its normalized label range is [-1, 1], so Tanh is used as the activation function.

[0120] Through the signal parameter recognition multi-task deep convolutional neural network can simultaneously regress and obtain the predicted values ​​of symbol rate and frequency offset.

[0121] S22. Set the loss function of the signal parameter recognition multi-task deep convolutional neural network to be the sum of two mean square errors multiplied by the corresponding weights, i.e., L mixed The network parameter update algorithm is the error back propagation algorithm (BP back propagation algorithm), the activation function corresponds to different regression tasks, set to Sigmoid and Tanh respectively, and the Adam optimizer is used for optimization. The initial learning rate lr is set to 0.0002, Adam∈ is set to 1e-8, Adamβ is set to (0.9, 0.999), weight decay is set to 0, and the lr update criterion is warmupcosinedecay;

[0122] Among them, L mixed It consists of two parts, namely the mean square error (MSE) of the two tasks multiplied by the corresponding weights:

[0123]

[0124] in, represents the mean square error of the fd task, represents the mean squared error of the offset task; μ1 and μ2 represent the weights of the mean squared errors of the fd and offset tasks, which are set to 0.3 and 0.7 during training. This loss function setting enables the two regression tasks to be trained together, which has a positive effect on model inference speed, reducing the impact of noise, and preventing overfitting of the subtasks.

[0125] The main parameters of the learning rate update criterion of Warmupcosine include: the number of warm-up training epochs warmup_epochs, the maximum learning rate base_value of the cosine learning rate curve, the minimum learning rate final_value, the period T_epochs of the cosine learning rate curve, and the decay of the cosine learning rate curve decay, that is, after each training epoch T_epochs, the maximum value of the learning rate curve will be multiplied by decay, and the number of iterations of each epoch n_iter.

[0126] The final learning curve is: at the beginning of the warm-up training, it increases linearly from 0 to the maximum value, then decreases from the maximum value to the minimum value in one cycle of cosine, and then jumps to the maximum value after decay, and so on until the training is completed.

[0127] Setting the learning rate curve in this way is conducive to the network model being trained to achieve good results in the warm-up training phase, and the jumps in the subsequent learning rate cycle changes can skip the local minimum point and continue training.

[0128] S23, shuffle the training samples and input them into the network for training;

[0129] The network is trained using two parallel multi-task labels, symbol rate fd and frequency offset offset, according to the loss function and optimization algorithm of step S22.

[0130] S24, testing the signal parameter recognition effect of the trained network;

[0131] The IQ data signals from the test sample set were input into the multi-task deep convolutional neural network for signal parameter identification. The mean absolute error (MAE) was used to measure the network's regression performance for symbol rate (fd) and frequency offset (offset), yielding parameter identification results. The test results for the first data set in S11 showed an average symbol rate (fd) error within 6.13% and an average frequency offset (offset) error within 1.74%. The symbol rate (fd) and frequency offset (offset) for the second data set in S11 fell within this range, ranging from -7% to 7% for fd and -2% to 2% for offset, respectively.

[0132] In this embodiment, step S3 is specifically as follows:

[0133] S31. Construct a deep convolutional neural network structure for signal modulation recognition based on the repvgg backbone structure;

[0134] Signal modulation recognition deep convolutional neural network structure such as Figure 3 As shown, its structure includes: input layer, backbone network layer, feature extraction layer, classification layer; and the network adopts the loss function binary cross entropy L CrossEntropyLoss Measure the error between the model prediction and the true target to optimize the model parameters; specifically as follows:

[0135] (1) Input layer: The input dimension of the network is (64, 2, 10000);

[0136] The first dimension represents the number of data in the data block of the data input, batch_size, which is set to 64 in this embodiment. The second dimension is the two IQ channels. The third dimension is the signal length of each IQ dimension.

[0137] (2) Backbone network layer: wherein the backbone network layer is the same as the backbone network of the multi-task deep convolutional neural network for signal parameter recognition in step S21;

[0138] (3) Feature extraction layer: The first layer of the feature extraction layer is the global average pooling layer GAP, and then the last dimension of the average pooling is removed. It passes through four consecutive layers of fully connected linear layers, the first three layers of rectified linear units ReLU, and the first three layers are all set with a dropout layer Dropout, with the dropout parameter set to 0.15, and finally outputs the feature vector of the number of categories;

[0139] The feature vectors obtained by the backbone network are used for the regression tasks of symbol rate fd and signal frequency offset offset.

[0140] The feature extraction layer structures used in the above two tasks are the same; both adopt a three-layer fully connected structure, the first two layers have rectified linear units (ReLU), and the first layer is set with a dropout layer (Dropout), the dropout parameter is set to 0.15, and the final output is the feature vector of the number of categories.

[0141] (4) Classification layer: Modulation recognition is a multi-classification task, so Softmax is used as the activation function.

[0142] S32. Set the loss function of the signal modulation recognition deep convolutional neural network to binary cross entropy L CrossEntropyLoss The network parameter update algorithm is the error back propagation algorithm (BP back propagation algorithm), the activation function is softmax, and the Adam optimizer is used for optimization. The initial learning rate lr is set to 0.0002, Adam∈ is set to 1e-8, Adamβ is set to (0.9, 0.999), the weight decay is set to, and the lr update criterion is warmupcosinedecay;

[0143] Loss function binary cross entropy L CrossEntropyLossThe expression is as follows:

[0144]

[0145] Where M represents the total number of signal modulation categories, N represents the number of samples in a batch; w m represents the weight of each category, x n,i Indicates the value of the nth data in the i-th label dimension, y n,m Represents the one-hot encoded label vector, which represents the value of the i-th category dimension of the n-th data label. If category x n is category m, then y n,m is 1, and other i not equal to m are 0.

[0146] The learning rate update criterion warmupcosine is the same as that introduced in step S22.

[0147] S33, training on the first data set described in step S11 to obtain a pre-trained model;

[0148] S34, fine-tuning the pre-trained model on the dataset with the parameter range narrowed down according to the result of parameter identification in step S24, i.e., the second dataset in step S11;

[0149] Use the fine-tune strategy of DeiT (Data-Efficient Image Transformer) in the image field:

[0150] Fully trained on a 224×224 resolution image dataset, and then fine-tuned on a high-resolution 384×384 dataset, the proposed method achieved similar or even better results than retraining on the 384×384 resolution data. Regarding signals, the dataset used in this paper is time-domain data. When the sampling rate is fixed, the lower the symbol rate, the more sampling points per signal information, i.e., the higher the resolution. Therefore, the pre-trained deep convolutional neural network model for signal modulation recognition is fine-tuned on a dataset with a low symbol rate.

[0151] In this embodiment, step S4 is specifically as follows:

[0152] The IQ data signal of the test sample set is input into the signal modulation recognition deep convolutional neural network. The position of the maximum result in the softmax output label dimension is used as the output modulation category result, and the accuracy is used to measure the network modulation recognition performance.

[0153] This embodiment also tests the test effect of the pre-trained model trained on the original data, and fine-tunes the small range parameter data using the pre-trained model after post-processing the data; the modulation recognition accuracy of the test changes with EsN as shown in Figure 4 The final comprehensive average accuracy rates are 78.25% and 91.81% respectively. It can be seen that the method of the present invention makes full use of the information of parameter identification in the modulation recognition task.

[0154] At the same time, this embodiment tests the effect of retraining other network structures on a dataset with a small parameter range and compares it with fine-tuning. Figure 5 As shown, the dataset includes ResNet50, a stacked Inception convolutional neural network, RepVGG, CaiT (Class-Attention in Image Transformers), and VitforSmall dataset. It can be seen that the final fine-tuning results of our method for this task are comparable to those of fresh training, while significantly reducing training time.

[0155] The above simulation experiments show that the method of the present invention can make good use of the information of parameter identification to further improve the effect of modulation identification; while achieving an accuracy comparable to that of retraining the modulation identification network, it reduces the consumption of computing resources.

[0156] In summary, the method of the present invention generates a modulation recognition data set by simulating the structure of an analog digital communication link, inputs the data set into a signal parameter recognition multi-task deep convolutional neural network with a repvgg as the main structure, and regresses to obtain the symbol rate and frequency deviation parameters of the signal. The obtained parameters can then be used to resample and frequency-shift the signal to obtain a small parameter range data set with a small symbol rate range and a small frequency deviation range, and then modulation recognition is performed on the data within the accuracy range of parameter recognition, that is, the small parameter range data is input into a modulation recognition pre-training model for fine-tune. The modulation recognition network has a repvgg as the main structure and is connected to a fully connected classification network. The pre-training model is generated by training the original data set, and then fine-tuned on a data set with a smaller parameter range to obtain a modulation category recognition model. The method of the present invention has a faster reasoning speed after using RepVGG as the backbone structure for deployment. It uses a deep convolutional neural network to identify signal parameters and fully utilizes the results for further training and processing. This enables the modulation recognition task to fully utilize the parameter information of the signal rather than relying solely on the waveform information of the signal for feature extraction. In addition, the use of fine-tune can maintain a signal modulation recognition effect similar to or even better than retraining in the downlink modulation recognition task with less training time.

[0157] Those skilled in the art will appreciate that the embodiments described herein are intended to aid the reader in understanding the principles of the present invention, and it should be understood that the scope of the present invention is not limited to such specific descriptions and embodiments. Various modifications and variations are readily apparent to those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims.

Claims

1. A modulation recognition optimization method for communication signals based on signal parameter estimation, the specific steps are as follows: S1. Build a training dataset; S2. Build and train a multi-task deep convolutional neural network for signal parameter recognition; S3. Build and train a deep convolutional neural network for signal modulation recognition to obtain a pre-trained model, and then fine-tune it within a small parameter dataset. S4, using the test sample set to obtain the modulation recognition accuracy after parameter estimation optimization; The step S1 is specifically as follows: S11. Use MATLAB simulation software to simulate a general communication link structure and generate a data set for signal parameter identification and modulation identification; Two datasets for signal parameter identification and modulation identification were generated in the same way: the first dataset was used for signal parameter estimation and modulation identification, while the second dataset was used only to fine-tune the pre-trained deep convolutional neural network model for signal modulation identification. The two datasets differed only in the symbol rate range and frequency offset range of the signal. in, The datasets used to generate the pre-trained models for the signal parameter estimation task and the signal modulation recognition task are the same, that is, the first dataset. The signal parameter estimation task uses the signal symbol rate and frequency offset as labels, while the signal modulation recognition task uses the modulation category as the label. S12, generates signals of 12 modulation types; In step S11, the data volume and channel conditions of the two data sets are the same, and the only difference is the range of symbol rate and the range of frequency offset; Each modulated signal contains 20,000 data items with a signal-to-noise ratio of 0-20dB, for a total of 240,000 data items; Among them, 80% are used as training sample sets, 10% are used as validation sample sets, and 10% are used as test sample sets. The symbol rate of the first data set varies randomly in the range of 100kHz to 200kHz, showing a uniform distribution, and the frequency offset varies randomly in the range of -10% to 10%, showing a uniform distribution. The symbol rate of the second data set varies randomly in the range of 93kHz to 107kHz, showing a uniform distribution, and the frequency offset varies randomly in the range of -2% to 2%, showing a uniform distribution. Set the channel conditions for generating data to Gaussian channel and Rayleigh channel; finally, store the IQ values ​​of the signal as a data set; S13, filtering the two data sets in step S11 through a low-pass filter and normalizing them; The data set is filtered through a Chebyshev low-pass filter with a set cutoff frequency, and then normalized to its maximum and minimum values. The normalization formula is as follows: Among them, Y i represents the i-th signal sample X i The sample value after normalization processing, Max represents the maximum value of a single sample data, and Min represents the minimum value of a single sample data; the signal labels include: modulation type, symbol rate, and frequency offset; The step S2 is specifically as follows: S21. Construct a multi-task deep convolutional neural network for signal parameter recognition based on the repvgg backbone structure; Signal parameter recognition multi-task deep convolutional neural network, its structure includes: input layer, backbone network layer, feature extraction layer, classification layer; and the network adopts loss function L mixed Measure the error between the model prediction and the true target and optimize the model parameters; specifically as follows: (1) Input layer: The input dimension of the network is (64, 2, 10000); The first dimension represents the number of data in the data block of the data input, batch_size; the second dimension represents the two IQ channels; and the third dimension represents the signal length of each IQ dimension. (2) Backbone network layer: adopts the repvgg block stacking structure and outputs the feature vector through the global average pooling layer GAP; The repvgg block mainly consists of two parts: Part 1: After the data is input, there are three branches, and then the three branches are superimposed together. The first branch consists of a convolution layer and a batch normalization layer, the second branch consists of a 1×1 convolution and a BN layer, and the third branch directly passes the data through the BN layer to the next layer; after the three branches are superimposed, they pass through the rectified linear unit ReLU; and only the first branch is multiplied by the learnable channel weight matrix M weight ; The working principle of this weight is expressed as follows: Among them, λ l,C The subscript C represents the number of channels of the first branch data input, and l represents the current depth of the repvgg block, that is, the network is composed of D repvgg blocks stacked together; the network structure expression of the first part is as follows: Where x′ l Represents the output of the first three branches in the lth repvgg block; The convolutional neural network representing the first branch of the first part of the repvgg block, x l Represents the input of the first part of the repvgg block and serves as the third branch of this part; Part 2: Input the superimposed features into a data channel adaptive weight block (SE-Block), that is, multiply each channel of the input data features by the weight value obtained by a channel weight extraction layer; The channel weight extraction layer is specifically composed of: a global average pooling layer GAP, a 1×1 convolution layer, a rectified linear unit ReLU, and a 1×1 convolution layer; The above two parts are combined and stacked to form a backbone network; (3) Feature extraction layer: The feature vector obtained by the backbone network is split into two channels, which are used for the regression tasks of symbol rate fd and signal frequency offset offset respectively; The feature extraction layer structure used in the above two tasks is the same, both adopt a three-layer fully connected structure. The first two layers have rectified linear units (ReLUs), and the first layer is set with a dropout layer with a dropout parameter of 0.

15. Finally, the feature vectors used for signal parameter regression are output respectively. (4) Classification layer: corresponds to two parallel feature extraction networks, and the activation function of the final output layer changes according to the regression parameters; Among them, for the symbol rate fd: its normalized label range is [0,1], so sigmoid is used as the activation function; for the frequency offset offset: its normalized label range is [-1,1], so Tanh is used as the activation function; The multi-task deep convolutional neural network can simultaneously regress the predicted values ​​of symbol rate and frequency offset through the signal parameter recognition; S22. Set the loss function of the signal parameter recognition multi-task deep convolutional neural network to be the sum of two mean square errors multiplied by the corresponding weights, i.e., L mixed The network parameter update algorithm is the error back propagation algorithm. The activation function corresponds to different regression tasks and is set to Sigmoid and Tanh respectively. The Adam optimizer is used for optimization. The initial learning rate lr is set to 0.0002, Adam∈ is set to 1e-8, Adamβ is set to (0.9, 0.999), weight decay is set to 0, and the lr update criterion is warmupcosinedecay. Among them, L mixed It consists of two parts, namely the mean square error of the two tasks multiplied by the corresponding weights: in, represents the mean square error of the fd task, represents the mean square error of the offset task; μ1 and μ2 represent the weights of the mean square error of the fd task and the offset task, which are set to 0.3 and 0.7 during training; The main parameters of the learning rate update criterion of Warmupcosine include: the number of warmup training epochs warmup_epochs, the maximum learning rate base_value of the cosine learning rate curve, the minimum learning rate final_value, the period T_epochs of the cosine learning rate curve, the decay of the cosine learning rate curve decay, that is, after each training epoch T_epochs, the maximum value of the learning rate curve will be multiplied by decay, and the number of iterations of each epoch n_iter; The final learning curve is: at the beginning of the warm-up training, it increases linearly from 0 to the maximum value, then decreases from the maximum value to the minimum value in one cycle, and then jumps to the maximum value after decay, and repeats this process until the training is completed; S23, shuffle the training samples and input them into the network for training; Using two parallel multi-task labels, symbol rate fd and frequency offset offset, and training the network according to the loss function and optimization algorithm of step S22; S24, testing the signal parameter recognition effect of the trained network; Input the IQ data signal of the test sample set into the signal parameter recognition multi-task deep convolutional neural network, use the mean absolute error (MAE) to measure the regression performance of the network's symbol rate fd and frequency offset offset, and obtain the parameter recognition result; The step S3 is specifically as follows: S31. Construct a deep convolutional neural network structure for signal modulation recognition based on the repvgg backbone structure; Construct a deep convolutional neural network structure for signal modulation recognition based on the repvgg backbone structure, which includes: input layer, backbone network layer, feature extraction layer, classification layer; and the network adopts the loss function binary cross entropy L CrossEntropyLoss Measure the error between the model prediction and the true target and optimize the model parameters; specifically as follows: (1) Input layer: The input dimension of the network is (64, 2, 10000); The first dimension represents the number of data in the data block of the data input, batch_size; the second dimension is the two IQ channels; and the third dimension is the signal length of each IQ dimension. (2) Backbone network layer: the same as the backbone network of the multi-task deep convolutional neural network for signal parameter recognition in step S21; (3) Feature extraction layer: The first layer of the feature extraction layer is the global average pooling layer GAP, and then the last dimension of the average pooling is removed. It passes through four consecutive layers of fully connected linear layers, the first three layers of rectified linear units ReLU, and the first three layers are all set with a dropout layer Dropout, with the dropout parameter set to 0.15, and finally outputs the feature vector of the number of categories; The feature vectors obtained by the backbone network are used for the regression tasks of symbol rate fd and signal frequency offset offset respectively; The feature extraction layer structure used in the above two tasks is the same; both use a three-layer fully connected structure, with the first two layers having rectified linear units (ReLUs), and the first layer is set with a dropout layer with a dropout parameter of 0.

15. The final output is a feature vector of the number of categories. (4) Classification layer: Modulation recognition is a multi-classification task, using Softmax as the activation function; S32. Set the loss function of the signal modulation recognition deep convolutional neural network to binary cross entropy L CrossEntropyLoss The network parameter update algorithm is the error back propagation algorithm, the activation function is softmax, and the Adam optimizer is used for optimization. The initial learning rate lr is set to 0.0002, Adam∈ is set to 1e-8, Adamβ is set to (0.9, 0.999), weight decay is set to 0, and the lr update criterion is warmupcosinedecay; Loss function binary cross entropy L CrossEntropyLoss The expression is as follows: Where M represents the total number of signal modulation categories, N represents the number of samples in a batch; w m represents the weight of each category, x n,i Indicates the value of the nth data in the i-th label dimension, y n,m Represents the one-hot encoded label vector, which represents the value of the i-th category dimension of the n-th data label. If category x n is category m, then y n,m is 1, and other i not equal to m are 0; The learning rate update criterion warmupcosine is the same as in step S22; S33, training on the first data set described in step S11 to obtain a pre-trained model; S34, fine-tuning the pre-trained model on the dataset with the parameter range narrowed down according to the result of parameter identification in step S24, i.e., the second dataset in step S11; Use DeiT's fine-tune strategy in the image field: The algorithm is fully trained on a 224×224 resolution image dataset and then fine-tuned on a high-resolution 384×384 dataset, achieving results similar to or even better than retraining on 384×384 resolution data.

2. The modulation identification optimization method of a communication signal based on signal parameter estimation according to claim 1, characterized in that: The step S4 is specifically as follows: The IQ data signal of the test sample set is input into the signal modulation recognition deep convolutional neural network, the position of the maximum result in the softmax output label dimension is used as the output modulation category result, and the accuracy is used to measure the network modulation recognition performance.

Images