Method for forming data set matched with multi-domain distribution under signal modulation identification task
By combining time domain and frequency domain information in the signal modulation identification task, using deep learning models to calculate distribution matching losses and update the synthetic data set, the problem that the data set synthesis method in the prior art failed to utilize multi-domain information is solved, and the data set synthesis method in the existing technology is realized. Efficient data set reduction and model performance improvement.
Patent Information
- Application Number
- CN202411190828.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-28
- Publication Date
- 2025-01-14
AI Technical Summary
The data set synthesis method in the existing signal modulation identification task fails to effectively combine information of different signal transformation domains, resulting in insufficient data set synthesis effect.
By calculating signal distribution matching losses in the time domain and frequency domain, combining with deep learning models to obtain low-dimensional embedding, updating the synthetic data set, using discrete Fourier transform to obtain frequency domain representation, and obtaining optimal synthesis by minimizing the objective function Dataset.
The models trained on smaller synthetic data sets have good classification capabilities, and the obtained synthetic data set has high compression rate, which improves the effect of data set synthesis.
Smart Images
Figure CN119312125A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of data set synthesis, and relates to a data set synthesis method matching multi-domain distribution under a signal modulation recognition task. Background Art
[0002] Modulation type recognition is a key step between signal detection and signal demodulation. Traditional modulation type recognition mainly relies on certain inherent features of the modulated signal to determine its modulation mode. However, there is a problem with traditional modulation methods, that is, they rely on the knowledge of domain experts to design artificial signal feature extraction models, which leads to insufficient feature representation capabilities. In recent years, deep learning technology has been gradually introduced into the electromagnetic signal modulation recognition task (reference [1]: Chen Z, Cui H, Xiang J, et al, "SigNet: A novel deep learning framework for radio signal classification," IEEE Transactions on Cognitive Communications and Networking, 2021, 8 (2): 529-541). Deep neural networks use a data-driven approach. Based on a large number of optimizable network parameters, they form a closed-loop structure through a back-propagation mechanism to achieve feature mining, which essentially solves the problem of insufficient traditional feature representation capabilities.
[0003] The success of deep learning is all due to training on large-scale datasets. However, such a large amount of data not only puts tremendous pressure on storage and transmission, but also brings great time and computing resource overhead to model training. One research direction to solve the large amount of data in deep learning is dataset synthesis, also known as dataset distillation. This field has been proposed in recent years and has been widely studied. Given an original large dataset, dataset synthesis aims to condense it and generate a small synthetic dataset so that the model trained on this small synthetic dataset performs similarly to the model trained on the original dataset. Some researchers have proposed a dataset distillation method based on representative distribution matching (reference [2]: Zhao B, Bilen H, "Dataset condensation with distribution matching," Proceedings of the IEEE / CVF Winter Conference on Applications of Computer Vision. 2023: 6514-6523, i.e. Zhao B, Bilen H, "Dataset synthesis method based on distribution matching" Proceedings of the IEEE / CVF Winter Conference on Applications of Computer Vision. 2023: 6514-6523). This method has achieved good experimental results in the image field, that is, it can still train a model with good classification ability on a small synthetic dataset. This guides us to conduct related research on dataset synthesis in the signal field.
[0004] Signals exist in multiple transform domains, such as time domain and frequency domain. Since different transform domains of signals contain rich and complementary information, combining information from different transform domains of signals has gradually become a trend. In recent years, methods that combine information from multiple transform domains of signals have begun to be applied in the field of signals (reference [3]: Peihan Qi, Xiaoyu Zhou, Shilian Zheng, and Zan Li, “Automatic Modulation Classification Based on Deep Residual Networks With Multimodal Information,” IEEE Transactions and Cognitive Communication and Network, Vol. 7, no. 1, Mar. 2021). However, there has been no research on combining information from different transform domains of signals for data set synthesis. Summary of the invention
[0005] In order to overcome the deficiency that the existing data set synthesis method under the signal modulation recognition task does not combine the information of different signal transform domains to improve the effect of data set synthesis, and to improve the effect of data set synthesis in the signal field, the present invention provides a data set synthesis method for matching multi-domain distribution under the signal modulation recognition task, which utilizes the time domain and frequency domain information of the modulated signal and combines the deep learning model to obtain the low-dimensional embedding of different data sets. The method updates the synthetic data set by matching the distribution of two data sets in the time domain and frequency domain respectively. A model with good classification ability can still be trained under a smaller synthetic data set. The synthetic data set finally obtained by the method has a high compression rate, and the model trained under the data set still has a high classification ability.
[0006] The technical solution adopted by the present invention to solve the technical problem is as follows:
[0007] A method for synthesizing a data set matching multi-domain distribution in a signal modulation recognition task comprises the following steps:
[0008] S1: Define the training set and synthetic dataset;
[0009] S2: Calculate the distribution matching loss between the training set signal and the synthetic set signal;
[0010] S3: Map the signals in the training set and the synthetic set to the frequency domain;
[0011] S4: Calculate the distribution matching loss between the training set signal and the synthetic set signal in the frequency domain;
[0012] S5: Combine the two distribution matching losses obtained in S2 and S4 as the final objective function, and obtain the optimal synthetic dataset by minimizing the objective function.
[0013] Further, the step S1 includes the following sub-steps:
[0014] S1.1: Divide the training set by category. The formula is as follows:
[0015]
[0016] Where T i represents the training sample with label i, |T i | represents the number of training samples of the i-th category, C represents the number of modulation types, Indicates T i The j-th signal in , each signal is represented by an I channel and a Q channel, and each channel contains N sampling points;
[0017] S1.2: Randomly sample from each type of training samples to obtain synthetic samples, and concatenate the synthetic samples of each type to obtain a synthetic dataset, which can be expressed as:
[0018]
[0019] Where S i ~T i Indicates that from T i Random sampling is used as synthetic samples, S i represents the synthetic sample with label i, S i | represents the number of synthetic samples of the i-th category, Indicates S i The mth signal in , each signal is represented by an I channel and a Q channel, and S represents the synthetic data set obtained after splicing.
[0020] Furthermore, the process of step S2 is as follows:
[0021] Randomly initialize a neural network model ψ θ , using ψ θ Get the low-dimensional embedding of real data and synthetic data, use the maximum average difference to estimate the distance between the real data distribution and the synthetic data distribution, and use this distance as the loss function in the time domain, expressed as:
[0022]
[0023] Among them, loss TD Represents the loss in the time domain.
[0024] Furthermore, step S3 includes the following sub-steps:
[0025] S3.1: In order to align with the time domain signal, the I channel and Q channel of each signal in the training set are discrete Fourier transformed and concatenated to obtain their frequency domain representation, which is expressed as:
[0026]
[0027] Where k is the number of the frequency domain signal, ranging from 1 to N. The number of points of the frequency domain signal is also N. The input of the discrete Fourier transform is N discrete points, that is, the time domain signal, and the output is N discrete points, that is, the frequency domain signal. Each point of the frequency domain signal is represented by a complex number. Take the amplitude of the frequency domain signal. Indicates T i The frequency domain representation of the jth signal in ;
[0028] S3.2: Perform the same operation on each signal in the synthetic dataset, expressed as:
[0029]
[0030] in Indicates S i Frequency domain representation of the mth signal in .
[0031] Furthermore, the process of step S4 is as follows:
[0032] Using ψ θ Obtain low-dimensional embeddings of real and synthetic signals from the frequency domain, use the maximum average difference to estimate the distance between the real data distribution and the synthetic data distribution in the frequency domain, and use this distance as the loss function in the frequency domain, expressed as:
[0033]
[0034] Among them, loss FD Represents the loss in the frequency domain.
[0035] Furthermore, step S5 includes the following sub-steps:
[0036] S5.1: Combining the two losses to obtain the objective function is expressed as:
[0037] loss=loss TD +α·loss FD (7)
[0038] Where α represents the weight coefficient;
[0039] S5.2: Update the synthetic dataset S by gradient descent until the objective function converges and obtains the optimal synthetic dataset, which is expressed as:
[0040]
[0041] Where η represents the learning rate, represents the gradient of the objective function with respect to the synthetic dataset, S * Represents the optimal synthetic dataset.
[0042] The technical concept of the present invention is: obtain the frequency domain representation of the modulated signal through discrete Fourier transform, combine the deep learning model to obtain the low-dimensional embedding of the training set and the synthetic data set, calculate the distribution matching loss between the two data sets in the time domain and frequency domain respectively, and organically combine the two losses to obtain the final objective function, and update the synthetic data set by minimizing this objective function.
[0043] The beneficial effects of the present invention are as follows: combining the time domain and frequency domain information of the modulated signal makes up for the deficiency of using only a single transform domain information, while combining the deep learning model to obtain the low-dimensional embedding of the training set and the synthetic data set, calculating the distribution matching loss between the two data sets from the time domain and frequency domain respectively, and organically combining the two as the final objective function, which makes up for the deficiency of the current data set synthesis method that is only performed on a single transform domain. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 This is a block diagram of a data set synthesis method that matches multi-domain distribution for the signal modulation recognition task proposed in the present invention.
[0045] Figure 2 This is a framework diagram of the deep learning model proposed in the present invention. DETAILED DESCRIPTION
[0046] The present invention will be described in further detail below in conjunction with the accompanying drawings.
[0047] Reference Figure 1 and Figure 2 , a data set synthesis method for matching multi-domain distribution under a signal modulation recognition task, this embodiment takes RML2016.10a high signal-to-noise ratio data as an example, and includes the following steps:
[0048] Step 1: Divide the original data set into a training set and a test set, re-divide the training set by category, randomly sample from each category of training samples to obtain synthetic samples, and splice the synthetic samples of each category to obtain a synthetic data set;
[0049] In step 1, the operation process is as follows: Figure 1 ,First, the RML2016.10a high signal-to-noise ratio data is divided into training set and test set with a ,distribution ratio of 4:1. Then the training set is re-divided by category, and a fixed number of ,samples are randomly sampled from each category of training samples as synthetic ,samples. Here the number we choose is 10, 50, 100. Then the synthetic samples of each category are ,joined to obtain the synthetic data set.
[0050] Step 2: Randomly initialize a deep learning model ψ θ , and input the training data and synthetic data of each class into the model to obtain their low-dimensional embeddings, calculate the maximum average difference between the two low-dimensional embeddings, and use it as the loss function in the time domain;
[0051] In step 2, the operation process is as follows: Figure 1 and attached Figure 2 We randomly initialize a deep learning model with seven layers, of which the first, third, and fifth layers are convolutional layers, the second, fourth, and sixth layers are maximum pooling layers, and the seventh layer is a fully connected layer. We input the training set and synthetic dataset signals obtained in step 1, obtain the output of the sixth layer of the model, calculate the maximum average difference between the two outputs, and use the difference as the loss function in the time domain.
[0052] Step 3: Perform discrete Fourier transform on the I channel and Q channel of each signal in each type of training samples and each type of synthetic samples, and then concatenate them to obtain the frequency domain representation of the signal; in step 3, the operation process is as follows: Figure 1 ,Each sample of the modulated signal dataset we use consists of an I channel and a Q channel, ,and each channel contains 128 sampling points. For each class of training samples and synthetic samples, ,the shape of each signal data is 1×2×128. Then we perform discrete Fourier transform on the I channel and Q channel of each signal, ,and take the amplitude, and then concatenate them, ,and the resulting frequency domain data has a shape of 1×2×128.
[0053] Step 4: Input the frequency domain representation of each type of training data and synthetic data into the model to obtain their low-dimensional embeddings, then use the maximum average difference to estimate the distance between the two low-dimensional embeddings, and use this distance as the loss function in the frequency domain;
[0054] In step 4, the specific operation process is as follows: the frequency domain representation of each type of training data and synthetic data obtained in step 3 is input into the model in step 2, the output of the sixth layer of the model is obtained, the maximum average difference between the two outputs is calculated, and the difference is used as the loss function in the frequency domain.
[0055] Step 5: Perform a weighted combination of the two loss functions obtained in step 2 and step 4 to obtain the final objective function, update the synthetic data set by the gradient descent method, set the number of update rounds, and use the synthetic data set obtained after the update as the final synthetic data set;
[0056] In step 5, the specific operation process is as follows: the loss functions obtained in steps 2 and 4 are weightedly combined to obtain the objective function, wherein when the number of synthetic samples in each category is 10 and 50, the optimal weight coefficient is 0.5, and when the number of synthetic samples in each category is 100, the optimal weight coefficient is 0.4, and then the synthetic data set is updated by minimizing the objective function, wherein the optimizer used to update the synthetic data set is the SGD optimizer, the momentum is set to 0.95, the weight decay is set to 0.00001, when the number of synthetic samples in each category is 10, the learning rate is 0.00001, when the number of synthetic samples in each category is 50 and 100, the learning rate is 0.0001, and the number of iterations is set to 20,000 rounds.
[0057] As described above, the present invention uses information from two different signal domains, time domain and frequency domain, to synthesize data sets in the case of using a modulated signal data set. The present invention defines a training set and a synthesis set, and then uses a randomly initialized deep learning model to obtain low-dimensional embeddings of the training set and the synthesis set, and estimates the distance between the two low-dimensional embeddings by the maximum average difference, which is used as a loss function in the time domain, and then obtains the frequency domain representation of the training set signal and the synthesis set signal by discrete Fourier transform, and then uses the previously randomly initialized deep learning model to obtain the low-dimensional embedding of the training set frequency domain signal and the synthesis set frequency domain signal, and calculates the maximum average difference between the two low-dimensional embeddings as the loss function in the frequency domain, and finally combines the loss functions in the time domain and the frequency domain as the final objective function, and sets a certain update round, and updates the synthesized data by the gradient descent method until the update is completed, and the updated synthesized data set is used as the final synthesized data set. It is only illustrative and not restrictive for the invention. It is understood by those skilled in the art that many changes, modifications, and even equivalents can be made to it within the spirit and scope defined by the claims of the invention, but they will all fall within the scope of protection of the present invention.
Claims
1. A method for synthesizing data sets matching multi-domain distribution in a signal modulation recognition task, characterized in that: The method comprises the following steps: S1: Define the training set and synthetic dataset; S2: Calculate the distribution matching loss between the training set signal and the synthetic set signal; S3: Map the signals in the training set and the synthetic set to the frequency domain; S4: Calculate the distribution matching loss between the training set signal and the synthetic set signal in the frequency domain; S5: Combine the two distribution matching losses obtained in S2 and S4 as the final objective function, and obtain the optimal synthetic dataset by minimizing the objective function.
2. The method for synthesizing a data set matching multi-domain distribution in a signal modulation recognition task according to claim 1, characterized in that: The step S1 comprises the following sub-steps: S1.1: Divide the training set by category. The formula is as follows: Where T i represents the training sample with label i, |T i | represents the number of training samples of the i-th category, C represents the number of modulation types, Indicates T i The j-th signal in , each signal is represented by an I channel and a Q channel, and each channel contains N sampling points; S1.2: Randomly sample from each type of training samples to obtain synthetic samples, and concatenate the synthetic samples of each type to obtain a synthetic dataset, which can be expressed as: Where S i ~T i Indicates that from T i Random sampling is used as synthetic samples, S i represents the synthetic sample with label i, S i | represents the number of synthetic samples of the i-th category, Indicates S i The mth signal in , each signal is represented by an I channel and a Q channel, and S represents the synthetic data set obtained after splicing.
3. The method for synthesizing a data set matching multi-domain distribution in a signal modulation recognition task according to claim 1 or 2, characterized in that: The process of step S2 is as follows: Randomly initialize a neural network model ψ θ , using ψ θ Get the low-dimensional embedding of real data and synthetic data, use the maximum average difference to estimate the distance between the real data distribution and the synthetic data distribution, and use this distance as the loss function in the time domain, expressed as: Among them, loss TD Represents the loss in the time domain.
4. The method for synthesizing a data set matching multi-domain distribution in a signal modulation recognition task according to claim 1 or 2, characterized in that: The step S3 comprises the following sub-steps: S3.1: In order to align with the time domain signal, the I channel and Q channel of each signal in the training set are discrete Fourier transformed and concatenated to obtain their frequency domain representation, which is expressed as: Where k is the number of the frequency domain signal, ranging from 1 to N. The number of points of the frequency domain signal is also N. The input of the discrete Fourier transform is N discrete points, that is, the time domain signal, and the output is N discrete points, that is, the frequency domain signal. Each point of the frequency domain signal is represented by a complex number. Take the amplitude of the frequency domain signal. Indicates T i The frequency domain representation of the jth signal in ; S3.2: Perform the same operation on each signal in the synthetic dataset, expressed as: in Indicates S i Frequency domain representation of the mth signal in .
5. The method for synthesizing a data set matching multi-domain distribution in a signal modulation recognition task according to claim 1 or 2, characterized in that: The process of step S4 is as follows: Using ψ θ Obtain low-dimensional embeddings of real and synthetic signals from the frequency domain, use the maximum average difference to estimate the distance between the real data distribution and the synthetic data distribution in the frequency domain, and use this distance as the loss function in the frequency domain, expressed as: Among them, loss FD Represents the loss in the frequency domain.
6. The method for synthesizing a data set matching multi-domain distribution in a signal modulation recognition task according to claim 1 or 2, characterized in that: The step S5 comprises the following sub-steps: S5.1: Combining the two losses to obtain the objective function is expressed as: loss=loss TD +α·loss FD (7) Where α represents the weight coefficient; S5.2: Update the synthetic dataset S by gradient descent until the objective function converges and obtains the optimal synthetic dataset, which is expressed as: Where η represents the learning rate, represents the gradient of the objective function with respect to the synthetic dataset, S * Represents the optimal synthetic dataset.