Communication signal generation method and device based on diffusion model

Abstract

The invention provides a communication signal generation method and device based on a diffusion model. The method comprises the following steps: randomly sampling from standard normal distribution to obtain a noise signal; obtaining the type of the target signal; and inputting the sampled noise signal and the target modulation type into a pre-trained diffusion model to generate a structured communication signal of the target modulation type. Compared with a communication signal generation mode based on a traditional method, complex model modeling is not needed, fine features and attributes of the communication signal are learned on the premise of no signal priori knowledge, and a high-fidelity communication signal of a specified modulation type is generated; compared with other generation methods based on deep learning, the method provided by the invention has better generalization performance, can guarantee stable training when learning high-dimensional data and complex signal distribution, is not liable to cause overfitting and mode collapse, and has higher interpretability and reliable theoretical derivation.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and in particular to a method and device for generating a communication signal based on a diffusion model. Background Art

[0002] With the development of emerging communication technologies such as 5G, Internet, satellite communication and physical layer security technology, modern communication systems are facing challenges such as limited spectrum resources, multi-user interference, environmental noise and multipath effects. As a result, existing communication signal generation methods can no longer meet the requirements of new communication systems for complex scenarios.

[0003] In the field of communications, signal generation needs include simulating real communication scenarios, generating signal data for testing and simulation, enhancing signal data sets to improve the training effect of machine learning models, etc. However, existing communication signal generation technologies have problems such as reliance on expert experience, poor generation quality, and poor interpretability. Therefore, the research on communication signal generation technology plays a vital role in the design, testing, simulation, and optimization of modern communication systems.

[0004] Existing signal generation methods can be roughly divided into signal generation methods based on traditional methods and signal generation methods based on deep learning. The former includes generation methods based on statistical models and generation methods based on physical models. According to specific signal characteristics such as frequency, amplitude, phase, etc., the signal is simulated by modeling generation process, and high-quality deterministic signals can be generated. The latter learns the complex probability distribution within the data through generative models such as generative adversarial networks and variational autoencoders to generate signals similar to the training data.

[0005] However, signal generation methods based on traditional methods usually rely on strict assumptions about the statistical characteristics of the signal or the generation process, and require certain prior knowledge. However, real-world signals, especially radio signals, are often highly complex and random, making them difficult to accurately model and generate using traditional methods. These traditional methods often perform poorly when faced with problems such as high noise environments, multipath effects, and complex modulation methods, and it is difficult to generate high-quality signals that are consistent with actual environments. Existing signal generation methods based on deep learning face a series of challenges, including unstable training processes, prone to mode collapse, and poor model interpretability. Summary of the invention

[0006] In order to solve the above problems existing in the prior art, the present invention provides a communication signal generation method and device based on a diffusion model, which specifically include:

[0007] In a first aspect, the present invention provides a communication signal generation method based on a diffusion model, comprising:

[0008] A noise signal is obtained by random sampling from a standard normal distribution;

[0009] Get the type of target signal;

[0010] The sampled noise signal and target modulation type are input into a pre-trained diffusion model to generate a structured communication signal of the target modulation type.

[0011] In a second aspect, the present invention further provides a communication signal generating device based on a diffusion model, comprising:

[0012] The acquisition module randomly samples a noise signal from a standard normal distribution;

[0013] The acquisition module is also used to obtain the type of the target signal;

[0014] The generation module is used to input the sampled noise signal and the target modulation type into a pre-trained diffusion model to generate a structured communication signal of the target modulation type.

[0015] In a third aspect, the present invention further provides an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0016] Memory, used to store computer programs;

[0017] The processor is used to implement any method provided in the first aspect when executing a program stored in the memory.

[0018] In a fourth aspect, the present invention provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, any one of the methods provided in the first aspect is implemented.

[0019] In a fifth aspect, the present invention provides a program product, which includes computer program instructions, and when the computer program instructions are executed, they can implement any of the methods provided in the first aspect.

[0020] Beneficial effects of the present invention:

[0021] The communication signal generation method and device based on the diffusion model provided by the present invention obtain a noise signal by random sampling from a standard normal distribution; obtain the type of the target signal; input the sampled noise signal and the target modulation type into a pre-trained diffusion model to generate a structured communication signal of the target modulation type. Compared with the communication signal generation method based on the traditional method, it does not require complex model building, and under the premise of no prior knowledge of the signal, such as the frequency, amplitude and other information of the signal, it learns the subtle features and properties of the communication signal to generate a high-fidelity communication signal of the specified modulation type; compared with other generation methods based on deep learning, the present invention has better generalization performance, and can ensure stable training when learning high-dimensional data and complex signal distribution, and is not prone to overfitting and mode collapse; in addition, compared with the black box characteristics of other deep learning methods, the forward diffusion process and signal generation process of the present invention have higher interpretability and reliable theoretical derivation.

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A flow chart of a communication signal generation method based on a diffusion model provided by the present invention;

[0024] Figure 2 An example of a signal generation process provided by the present invention;

[0025] Figure 3 The present invention provides a MLP label Structural diagram;

[0026] Figure 4 The present invention provides a structure of a communication signal generating device based on a diffusion model. DETAILED DESCRIPTION

[0027] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0028] The diffusion model gradually generates signals by adding noise to the forward diffusion of the communication signal and learning to remove noise in reverse. This slow, probability-based learning process enables the diffusion model to be stably trained, learn complex signal characteristics, and generate high-quality, detailed communication signals in a changeable and complex electromagnetic environment. Based on this, the present invention proposes a communication signal generation method based on a diffusion model, which can solve the problems of low generation quality and poor model interpretability of existing communication signal generation technologies.

[0029] Figure 1A flow chart of a communication signal generation method based on a diffusion model provided by the present invention is shown in FIG. Figure 1 As shown, the method includes:

[0030] S101. A noise signal is obtained by random sampling from a standard normal distribution.

[0031] Specifically, a noise signal s is randomly sampled from a standard normal distribution T , as the starting point of the generation process.

[0032] S102: Obtain the type of the target signal.

[0033] S103: Input the sampled noise signal and target modulation type into a pre-trained diffusion model to generate a structured communication signal of the target modulation type.

[0034] like Figure 2 As shown in the figure, specifically, starting from time T, the diffusion model gradually denoises according to the reverse process, and at each step, the diffusion model is used to predict the noise of the previous time step. The above denoising process is repeated continuously until the signal is generated from time step T to time step 0. In this process, the model continuously converts the noise into a structured communication signal.

[0035] Furthermore, for any training sample, in a possible implementation, training the diffusion model includes the following steps A1 and A2:

[0036] A1. Perform forward noise addition on the communication signal through the forward diffusion process. After diffusing the original signal s0 for any number of steps t, the noised signal s is obtained. t , expressed as:

[0037]

[0038] Among them, q(s t |s0) represents the intermediate state s in the diffusion process t The probability distribution under the initial state s0, that is, the data s at time step t when the initial data s0 is known t The distribution of Indicates t The distribution satisfies the mean The variance is Gaussian distribution, α t =1-β t , β t ∈[β min ,β max ] represents the Gaussian noise covariance parameter added in the tth step, I represents the identity matrix, T represents the total diffusion step length, s t(1≤t≤T) represents the signal obtained by diffusion in the tth step.

[0039] Specifically, in the forward diffusion process, noise is continuously added to the communication signal s0 that carries information, gradually eliminating the information carried by the signal, and finally obtaining a noise vector.

[0040] The process can be expressed as follows:

[0041]

[0042] Among them, the input signal s0 satisfies the distribution q(s0), T is the total diffusion step length, s t (1≤t≤T) represents the signal obtained by diffusion in the tth step, β t ∈[β min ,β max ] is the Gaussian noise covariance parameter added in the tth step. Each step of the noise addition process is derivable and controllable. Therefore, the communication signal s with any diffusion step length t It can be directly obtained from the initial input s0 and t:

[0043]

[0044] Among them, α t =1-β t , The entire forward diffusion process does not contain any trainable parameters.

[0045] For example, the signal s0 is in IQ form with a dimension of 2×128; the total diffusion step length T=1000; the minimum noise covariance β min =0.0001, maximum noise covariance β max =0.02.

[0046] A2, through the reverse diffusion process, according to the modulation type of the original signal s0 and the signal s after noise addition t , estimate the added noise, and complete the training of the diffusion model based on the estimation results and the preset loss function.

[0047] The inverse diffusion process continuously removes noise, restores information details, and ultimately generates a signal that carries information.

[0048] In order to improve the generation quality of the model, the category labels of the samples are used to guide the generation process of the diffusion model during the model training process.

[0049] Specifically, after the reverse diffusion process, according to the modulation type of the original signal s0 and the noise-added signal s t , the estimated added noise is expressed as:

[0050]

[0051] Cond=MLP time (t)+MLP label (y),

[0052] Among them, s t-1 represents the signal obtained by diffusion in the t-1th step, ε θ represents the network used to predict noise, Cond represents the guidance used in the diffusion model training process, represents a standard normal distribution with a mean of 0 and a variance of 1, and y represents the label of the target modulation type.

[0053] Specifically, the guidance used in the diffusion model training process is specifically expressed as:

[0054] Cond=MLP time (s0)+MLP label (y),

[0055] Among them, MLP time (s0) means mapping the original signal s0 into a vector of length 256, MLP time It is a one-dimensional convolution structure with 2 input channels, 64 output channels, 7 convolution kernels, and 3 padding.

[0056] MLP label (y) represents mapping the label y of the target modulation type to a modulation type constraint vector with a length of 256. Figure 3 As shown, MLP label It includes a cosine position encoding layer and two linear connection layers with an output dimension of 256.

[0057] Specifically, the cosine position coding layer is used to encode the label of the target modulation type to obtain a coding vector of dimension d, which is expressed as:

[0058]

[0059] Among them, pos represents the position in the sequence, i represents the index value of the dimension, PE(pos,2i) and PE(pos,2i+1) represent the position encoding in the even dimension and odd dimension respectively.

[0060] Exemplarily, d = 64. The coding vector with a dimension of 64 output by the cosine position coding layer passes through two layers of linear connection layers with an output dimension of 256 in sequence, and finally outputs a modulation type constraint vector with a dimension of 256.

[0061] Optionally, a loss function, expressed as:

[0062]

[0063] Among them, ||·|| represents the bi-norm of the vector, represents the loss function of the diffusion model, θ represents the model parameters, Ε represents the expectation, ε represents the noise term actually added, and ε θ represents the prediction noise, t represents the time step of the diffusion process, and y represents the true label of the signal.

[0064] Specifically, the back-propagation algorithm can be used to optimize the weights in the loss function and gradually adjust the model parameters to reduce the difference between the generated signal and the real signal. The AdamW optimizer is used to smooth and stabilize the parameter update process and improve the model robustness.

[0065] For example, the total number of training rounds is set to 5000. The learning rate is set to α = 0.002, the momentum decay hyperparameters are set to β1 = 0 and β2 = 0.999, the weight decay coefficient λ = 0, and the minimum constant ε to prevent division by zero = 1 × 10 -8 .

[0066] The present invention uses a diffusion model to ensure the quality of generated communication signals, improves the interpretability of a generation method based on deep learning, and guides the generation process of the model by mapping the modulation type of the communication signal into a coding vector through cosine position coding and a fully connected layer.

[0067] The communication signal generation method based on the diffusion model provided by the present invention obtains a noise signal by random sampling from a standard normal distribution; obtains the type of the target signal; inputs the sampled noise signal and the target modulation type into a pre-trained diffusion model to generate a structured communication signal of the target modulation type. Compared with the communication signal generation method based on the traditional method, it does not require complex model building, and under the premise of no prior knowledge of the signal, such as the frequency and amplitude of the signal, it learns the subtle features and properties of the communication signal to generate a high-fidelity communication signal of the specified modulation type; compared with other generation methods based on deep learning, the present invention has better generalization performance, and can ensure stable training when learning high-dimensional data and complex signal distribution, and is not prone to overfitting and mode collapse; in addition, compared with the black box characteristics of other deep learning methods, the forward diffusion process and signal generation process of the present invention have higher interpretability and reliable theoretical derivation.

[0068] Figure 4 A structural diagram of a communication signal generating device based on a diffusion model provided by the present invention, such as Figure 4 As shown, the device comprises:

[0069] The acquisition module 41 randomly samples from a standard normal distribution to obtain a noise signal.

[0070] The acquisition module 41 is further used to acquire the type of the target signal.

[0071] The generating module 42 is used to input the sampled noise signal and the target modulation type into a pre-trained diffusion model to generate a structured communication signal of the target modulation type.

[0072] In order to further demonstrate the beneficial effects of the present invention, the present invention also verifies the quality of the signal generated according to the present invention, as follows:

[0073] Based on the existing method, a communication signal classifier is obtained by training a training set, and the training set is consistent with the training set of the diffusion model in the present invention. After the model converges, the model parameters are saved. The classifier model parameters are loaded, and the communication signals in the test set are modulated and classified to obtain the final modulation classification accuracy.

[0074] The experimental results are shown in Table 1, which respectively gives the recognition accuracy of the classifier in the test set and the generated signal, as well as the difference between the two accuracy rates.

[0075] Table 1

[0076]

[0077] It can be seen from the results in Table 1 that when the signal ratio range of the communication signal is 0 to 18 dB, the difference in recognition accuracy of the classifier in the test set and the generated signal ranges from 0.32% to 1.37%, which shows that the generated signal has the characteristics of the specified modulation category. The method of the present invention learns the complex distribution characteristics of the signal, and the communication signal of the specified modulation type generated by it has the advantage of high quality.

[0078] The present invention also provides a structure of an electronic device, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus.

[0079] Memory, used to store computer programs;

[0080] The processor is used to implement the steps provided in the above method embodiment when executing the program stored in the memory.

[0081] The communication interface is used for communication between the above electronic device and other devices.

[0082] The method provided in the embodiment of the present invention can be applied to electronic devices. Specifically, the electronic device can be: a desktop computer, a portable computer, an intelligent mobile terminal, a server, etc. This is not limited here, and any electronic device that can implement the present invention belongs to the protection scope of the present invention.

[0083] The present invention also provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the steps provided in the above method embodiment are implemented.

[0084] The present invention provides a program product, which includes computer program instructions. When the computer program instructions are executed, the steps provided in the above method embodiment can be implemented.

[0085] As for the device / electronic device / storage medium / program product embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the specific contents and beneficial effects and other related matters can be referred to the partial description of the method embodiments.

[0086] The terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0087] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A communication signal generation method based on a diffusion model, characterized in that: include: A noise signal is obtained by random sampling from a standard normal distribution; Get the type of target signal; The sampled noise signal and the target modulation type are input into a pre-trained diffusion model to generate a structured communication signal of the target modulation type.

2. The method according to claim 1, characterized in that Training diffusion models, including: For any training sample, The forward diffusion process performs forward noise addition on the communication signal. After diffusing the original signal s0 for any number of steps t, the noised signal s is obtained. t , expressed as: Among them, q(s t |s0) represents the intermediate state s in the diffusion process t The probability distribution under the initial state s0 is: Indicates t The distribution satisfies the mean The variance is Gaussian distribution, α t =1-β t , β t ∈[β min ,β max ] represents the Gaussian noise covariance parameter added in the tth step, I represents the identity matrix, T represents the total diffusion step length, s t (1≤t≤T) represents the signal obtained by diffusion in the tth step; After the reverse diffusion process, according to the modulation type of the original signal s0 and the noise-added signal s t , estimate the added noise, and complete the training of the diffusion model based on the estimation results and the preset loss function.

3. The method according to claim 2, characterized in that The reverse diffusion process is performed according to the modulation type of the original signal s0 and the noise-added signal s t , the estimated added noise is expressed as: Cond=MLP time (t)+MLP label (y), Among them, s t-1 represents the signal obtained by diffusion in the t-1th step, ε θ represents the network used to predict noise, Cond represents the guidance used in the diffusion model training process, represents a standard normal distribution with a mean of 0 and a variance of 1, and y represents the label of the target modulation type.

4. The method according to claim 3, characterized in that The guidance used in the diffusion model training process is specifically expressed as: Cond=MLP time (s0)+MLP label (y), Among them, MLP time (s0) means mapping the original signal s0 into a vector of length 256, MLP time It is a one-dimensional convolution structure with 2 input channels, 64 output channels, 7 convolution kernels, and 3 padding. MLP label (y) represents mapping the target modulation type label y to a modulation type constraint vector of length 256, MLP label It includes a cosine position encoding layer and two linear connection layers with an output dimension of 256.

5. The method according to claim 4, characterized in that The cosine position coding layer is used to encode the label of the target modulation type to obtain a coding vector with a dimension of d, which is expressed as: Among them, pos represents the position in the sequence, i represents the index value of the dimension, PE(pos,2i) and PE(pos,2i+1) represent the position encoding in the even dimension and odd dimension respectively.

6. The method according to claim 5, characterized in that The loss function is expressed as: Among them, ||·|| represents the bi-norm of the vector, represents the loss function of the diffusion model, θ represents the model parameters, Ε represents the expectation, ε represents the noise term actually added, and ε θ represents the prediction noise, t represents the time step of the diffusion process, and y represents the true label of the signal.

7. A communication signal generating device based on a diffusion model, characterized in that: include: The acquisition module randomly samples a noise signal from a standard normal distribution; The acquisition module is further used to acquire the type of the target signal; The generating module is used to input the sampled noise signal and the target modulation type into a pre-trained diffusion model to generate a structured communication signal of the target modulation type.

8. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, for implementing any of the methods described in claims 1-6 when executing a program stored in a memory.

9. 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 6 is implemented.

10. A program product, characterized in that The program product comprises computer program instructions, and when the computer program instructions are executed, the method according to any one of claims 1 to 6 can be implemented.

Images