Conditional communication signal generation method based on de-noising diffusion probability model

Abstract

The invention discloses a conditional communication signal generation method based on a de-noising diffusion probability model. The method comprises a diffusion process and a de-noising process. According to the method, the signal generation process is guided and optimized by using specific condition information, so that the generated signal can better meet the requirements of specific application scenes, and high-quality, efficient and controllable communication signal generation is realized through an effective DDPM signal generation framework, a reasonable network architecture design and an efficient reasoning optimization strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal generation, in particular to a conditional communication signal generation method based on a denoising diffusion probabilistic model. BACKGROUND

[0002] Communication signal generation technology refers to a technology that uses algorithmic models to simulate the data distribution of real target communication signals and achieve high-quality approximation signal or waveform generation. Thanks to the rapid development of generative artificial intelligence technology (GAI), artificial intelligence models have the ability to learn signal underlying features from complex electromagnetic environments and electromagnetic big data, and on this basis, they can generate high-quality communication signals with strong generalization ability. Currently, artificial intelligence-based communication signal generation technology has played a crucial role in signal intelligence augmentation, fingerprint feature learning, and channel effect simulation.

[0003] Most current communication signal generation methods are based on generative adversarial networks (GANs). Due to the unbalanced competition between the generator and the discriminator in the GAN model, the training process of GAN is inherently unstable. During the training process, the discriminator often gains a decisive advantage, leading to the occurrence of mode collapse. In addition, the training of the GAN model is very sensitive to hyperparameters, including learning rate, batch size, and network architecture. Small differences in these hyperparameters can lead to completely different data generation results. The above problems pose a great challenge to achieving high-quality communication signal generation.

[0004] Although the diffusion probabilistic model (DDPM) has strong learning and generation capabilities for image and speech data, its application to the field of communication signal generation is still very challenging. Compared with speech data, communication signal data has different data characteristics and higher sampling rates, which makes the neural network architecture and reception field in DDPM unable to meet the needs of communication signal generation tasks. In addition, the inference of DDPM often requires thousands of iterations of the denoising process, which is very inefficient and difficult to meet the requirements of most communication signal generation tasks for algorithm efficiency. SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art and provide a conditional communication signal generation method based on a denoising diffusion probabilistic model. This method uses specific conditional information to guide and optimize the signal generation process, making the generated signals better meet the needs of specific application scenarios.

[0006] The technical scheme adopted by the present application is:

[0007] A conditional communication signal generation method based on a denoising diffusion probability model, including a diffusion process and a denoising process;

[0008] The diffusion process is specifically:

[0009] The diffusion process is defined as a T-step Markov chain, which gradually transforms the original communication signal x0 into a complete Gaussian noise xT. T :

[0010]

[0011] where q(x t |x t-1 ) adopts a pre-defined variance sequence, denoted as [β0,…,β T ], to introduce Gaussian noise into the distribution of x t-1 , which satisfies the following conditions:

[0012]

[0013] Let z t represent a Gaussian noise signal, then formula (2) can be expressed as:

[0014]

[0015] Define α t = 1-β t , then we get:

[0016]

[0017] Formula (4) can be recursively expanded as:

[0018]

[0019] where

[0020] The denoising process uses the Bayes formula to derive the inverse process q(x t |x t-1 ) from the diffusion process q(x t-1 |x t ),

[0021]

[0022] Specifically, use x0 to calculate the above formula:

[0023]

[0024] From equation (2) and equation (5), we can get:

[0025]

[0026] Through the above derivation, we can get the distribution of the inverse process q(x t-1 |x t ,x0):

[0027]

[0028] where, is a constant, for each diffusion step t, it is independent of x t , so μ θ (x t ,t) can be written as:

[0029]

[0030] where, the neural network model ε θ (x t ,t) is trained to get μ θ (x t ,t), and x0 is obtained by putting equation (9) into the formula, completing the iterative denoising of the input Gaussian noise signal x T .

[0031] Specifically, in the denoising process, a conditional label c is added to the neural network model ε θ (x t ,t) to control the output of the inverse process, and the training process is as follows:

[0032] Step 1, input signal data set X, iteration step T, conditional label set C, parameter

[0033] Step 2, randomly select a set of training data x0 from X;

[0034] Step 3, select the corresponding conditional label c from the conditional label set C;

[0035] Step 4, randomly select a step parameter t from {1…T};

[0036] Step 5, generate a set of noise signals ε ~ N(0,1);

[0037] Step 6, calculate x t according to the formula

[0038] Step 7, minimize the loss function L = || ε - ε θ (x t ,t,c) ||​2 , optimize the update network model epsilon θ (x t , t, c);

[0039] Step 8, repeat steps 2-7 until training converges, get the trained neural network model epsilon θ (x t , t, c)。

[0040] Due to the adoption of the technical solutions as described above, the present application has the following advantages:

[0041] The present application uses specific condition information to guide and optimize the signal generation process, so that the generated signal can better meet the needs of specific application scenarios. Through an effective DDPM signal generation framework, reasonable network architecture design and efficient inference optimization strategy, the present application realizes high-quality, efficient and controllable communication signal generation. BRIEF DESCRIPTION OF DRAWINGS

[0042] Fig. 1 is the basic principle of the communication signal generation based on the denoising diffusion probability model of the present application.

[0043] Fig. 2 is a single signal generation framework based on DDPM of the present application. DETAILED DESCRIPTION

[0044] The present application will be further explained and described below in conjunction with the drawings and embodiments, which cannot limit the protection scope of the present application. The purpose of disclosing the present application is to protect all technical improvements within the scope of the present application.

[0045] In conjunction with the drawings Figs. 1-2 A conditional communication signal generation method based on a denoising diffusion probability model, characterized in that it comprises a diffusion process and a denoising process.

[0046] The diffusion process is specifically:

[0047] The diffusion process is defined as a T-step Markov chain, which gradually transforms the original communication signal x0 into a complete Gaussian noise x T :

[0048]

[0049] Wherein, q(x t |x t-1 ) adopts a pre-defined variance sequence, denoted as [beta0, …, beta T ], which introduces Gaussian noise into the distribution of x t-1 , which satisfies the following conditions:

[0050]

[0051] Let z t be a Gaussian noise signal, then equation (2) can be expressed as:

[0052]

[0053] Let α t = 1-β t , then we have:

[0054]

[0055] Equation (4) can be recursively expanded as:

[0056]

[0057] where

[0058] The denoising process uses Bayes' theorem to derive the inverse process q(x t |x t-1 ) from the diffusion process q(x t-1 |x t ),

[0059]

[0060] Specifically, the above formula is calculated using x0:

[0061]

[0062] From equation (2) and equation (5), we have:

[0063]

[0064] Through the above derivation, the distribution of the inverse process q(x t-1 |x t ,x0) can be obtained:

[0065]

[0066] where, is a constant, for each diffusion step t, it is independent of x t , so μ θ (x t ,t) can be written as:

[0067]

[0068] where the neural network model ε θ (x t ,t) is trained to obtain μ θ(x t , t) into equation (9) to get x0, and the iteration denoising of input Gaussian noise signal x T is completed.

[0069] In the denoising process, a conditional label c is added in the neural network model ε θ (x t , t) to control the output of the inverse process, and the training process is specifically as follows:

[0070] Step 1, input signal data set X, iteration step length T, conditional label set C, and parameters

[0071] Step 2, randomly select a set of training data x0 from X;

[0072] Step 3, select the corresponding conditional label c from the conditional label set C;

[0073] Step 4, randomly select a step length parameter t from {1…T};

[0074] Step 5, generate a set of noise signals ε ~ N(0, 1);

[0075] Step 6, calculate x t according to the formula ;

[0076] Step 7, optimize and update the network model ε θ (x t , t, c) by minimizing the loss function L = ||ε-ε θ (x t , t, c) || 2 ;

[0077] Step 8, repeat steps 2-7 until the training converges, and obtain the trained neural network model ε θ (x t , t, c).

[0078] The parts of the application not described in detail are prior art.

[0079] The examples selected in the text for the purpose of disclosing the inventive object of the present application are currently considered appropriate, but it should be understood that the present application is intended to include all variations and improvements of the examples that belong to the scope of the present concept and invention.

Claims

1. A conditional communication signal generation method based on a denoising diffusion probability model, characterized in that: This includes the diffusion process and the denoising process; The diffusion process is specifically as follows: The diffusion process is defined as a T-step Markov chain that gradually transforms the original communication signal x0 into complete Gaussian noise x. T : Where q(x) t |x t-1 A predefined variance sequence is used, denoted as [β0,…,β]. T ], towards x t-1 When Gaussian noise is introduced into the distribution, it satisfies the following condition: Let z t Let Gaussian noise signal be represented, then formula (2) can be expressed as: Define α t =1-β t Then we get: Formula (4) can be recursively expanded as follows: in The denoising process uses Bayes' theorem, from the diffusion process q(x) t |x t-1 The inverse process q(x) is derived. t-1 |x t ), Specifically, x0 is used to calculate the above formula: From equations (2) and (5), we can conclude that: Through the above derivation, the inverse process q(x) can be obtained. t-1 |x t Distribution of x0): in, It is a constant that, for each diffusion step t, is related to x. t Irrelevant, therefore μ θ (x t ,t) can be written as: Among them, the neural network model ε θ (x t μ is obtained after training (t). θ (x t Substituting ,t), into formula (9) yields x0, thus completing the processing of the input Gaussian noise signal x. T Iterative denoising.

2. The conditional communication signal generation method based on a denoising diffusion probability model according to claim 1, characterized in that: In the aforementioned denoising process, in the neural network model ε θ (x t A conditional label c is added to (t) to control the output of the inverse process. The training process is as follows: Step 1, Input signal dataset X, iteration step size T, condition label set C, parameters Step 2: Randomly select a set of training data x0 from X; Step 3: Select the corresponding condition label c from the condition label set C; Step 4: Randomly select a step size parameter t from {1…T}; Step 5: Generate a set of noise signals ε ~ N(0,1); Step 6, according to the formula Calculate x t ; Step 7, by minimizing the loss function L = ||ε - ε θ (x t ,t,c)|| 2 Optimize and update the network model ε θ (x t ,t,c) Step 8: Repeat steps 2-7 until training converges, obtaining the trained neural network model ε. θ (x t ,t,c).

Images