A signal intelligent generation method and device based on diffusion model

Abstract

The present invention relates to a method and device for intelligently generating signals based on a diffusion model, which belongs to the technical field of communication signal management, and solves the problem that it is difficult for the prior art to effectively manage signals without prior knowledge or feedback information. The method of the present invention includes: preprocessing a target signal to obtain a preprocessed signal; adding noise to the preprocessed signal based on a diffusion process to obtain a noise injection signal; constructing a noise estimation model, and training the noise estimation model using the noise injection signal; selecting a pure noise signal, and inputting the pure noise signal into the trained noise estimation model to obtain a noise estimation value of the pure noise signal; based on the inverse process, denoising the pure noise signal according to the noise estimation value, thereby obtaining a generated signal similar to the target signal. The present invention can generate a signal similar to the target signal without prior knowledge or feedback information, thereby achieving management of dangerous signals.

Description

Technical Field

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

[0002] With the continuous application of new communication technologies and new information processing methods, more and more new devices are emerging, and their intelligence levels are getting higher and higher. While new device technologies and intelligent communication technologies bring convenience to society, they also bring dangers. How to effectively manage new and unknown signals is an urgent problem that needs to be solved.

[0003] In the field of communications, signal management is a key technical link, which involves multiple aspects such as signal generation, processing, transmission and reception. The main goal of signal management is to ensure the quality and integrity of the signal during transmission, and to accurately restore the original information at the receiving end. At present, the management of wireless signals usually requires obtaining feedback signals, or conducting preliminary research on the target signal to be managed, and obtaining some prior knowledge, such as the frequency band, modulation style and technical parameters of the signal. However, with the continuous emergence of new intelligent communication signal technologies, the acquisition and processing of feedback signals has become more difficult, and the prior knowledge of new signals is also difficult to obtain. In the prior art, for new wireless signals emitted by new devices, and other signals without prior knowledge or feedback information, more advanced technologies such as generative adversarial networks are usually used for management. However, generative adversarial network technology has problems such as gradient explosion and instability, and cannot achieve effective management of wireless signals. Summary of the invention

[0004] In view of the above analysis, the embodiments of the present invention aim to provide a method and device for intelligently generating signals based on a diffusion model, so as to solve the problem that it is difficult to effectively manage signals without prior knowledge or feedback information in the prior art.

[0005] On the one hand, an embodiment of the present invention provides a method and device for intelligently generating a signal based on a diffusion model, the method comprising:

[0006] Preprocessing a target signal without prior knowledge to obtain a preprocessed signal;

[0007] Adding noise to the preprocessed signal to obtain a noise injection signal;

[0008] Constructing a noise estimation model, and using the noise injection signal to train the noise estimation model;

[0009] Selecting a pure noise signal, and inputting the pure noise signal into a trained noise estimation model to obtain a noise estimation value of the pure noise signal;

[0010] The pure noise signal is denoised according to the noise estimation value, thereby obtaining a generated signal similar to the target signal.

[0011] Based on the further improvement of the above method, the adding noise to the preprocessed signal to obtain the noise injection signal includes:

[0012] The preprocessed signal is processed based on the diffusion process Processing, through Iterations add noise to the preprocessed signal , thereby obtaining the noise injection signal , is an integer greater than 1.

[0013] Based on the further improvement of the above method, the diffusion process includes a fixed Markov chain propagation process;

[0014] Among them, based on the fixed Markov chain propagation process, through Iterate over the preprocessed signal Add noise to , thereby obtaining the noise injection signal , the formula is:

[0015] ; ;

[0016] ;

[0017] ;

[0018] In the formula, is an integer greater than 1, , The preprocessed signals are After t times, The signal after adding noise in iterations is is the iteration parameter, express It conforms to the standard normal distribution, and Draw random values ​​from a standard normal distribution, is the variance of the standard normal distribution at the t-th iteration, Pick a random value from the set of variance values.

[0019] Based on the further improvement of the above method, the reverse process is a learnable Markov chain propagation process;

[0020] Wherein, based on the learnable Markov chain propagation process, the pure noise signal is Perform T steps of iteration to perform denoising processing, thereby obtaining a generated signal similar to the target signal , the formula is:

[0021] ; ;

[0022] ;

[0023] In the formula, , Pure noise signal go through , The signal after the denoising process is performed iteratively. for Input the signal noise estimate obtained from the trained noise estimation model, yes The cumulative mean of the standard normal distribution of iterations; express It conforms to the standard normal distribution, and hour, Draw random values ​​from a standard normal distribution; hour, .

[0024] Based on a further improvement of the above method, the pure noise signal is a random noise signal, and the signal value of the random noise signal is randomly taken from a standard normal distribution.

[0025] Based on the further improvement of the above method, the loss function of the noise estimation model is:

[0026] ;

[0027] ;

[0028] In the formula, is the noise loss, For signal The noise estimation value obtained by inputting the noise estimation model, Represents the L2 loss function operation.

[0029] Based on the further improvement of the above method, the noise estimation model is a TUnet network.

[0030] The TUnet network includes:

[0031] Residual network, used for preliminary feature extraction of input signals;

[0032] A downsampling network block, used to perform a downsampling operation on the data output by the residual network;

[0033] An intermediate processing network block, used for performing a deconvolution operation on the data output by the downsampling network block;

[0034] An upsampling network block, used for performing an upsampling operation on the data output by the intermediate processing network block and the data output by the downsampling network block;

[0035] A time information embedding network block is used to input time information into the upsampling network block, thereby embedding time information in output data of the upsampling network block and obtaining upsampling data embedded with time information;

[0036] The output network block is used to further extract features from the upsampled data embedded with the time information to generate a final output signal.

[0037] Based on the further improvement of the above method, the method further includes:

[0038] Performing kernel density estimation on the generated signal and the preprocessed signal respectively, and drawing corresponding probability density curves;

[0039] The probability density curves of the generated signal and the preprocessed signal are compared to determine whether the probability distributions of the two signals are the same, so as to evaluate the similarity between the generated signal and the target signal.

[0040] Based on the further improvement of the above method, the method further includes:

[0041] The KL divergence between the generated signal and the preprocessed signal is calculated to evaluate the similarity between the generated signal and the target signal.

[0042] On the other hand, an embodiment of the present invention provides a signal intelligent generation device based on a diffusion model, the device comprising:

[0043] A preprocessing module preprocesses the target signal without prior knowledge to obtain a preprocessed signal;

[0044] A noise injection module, adding noise to the preprocessed signal to obtain a noise injection signal;

[0045] A noise estimation model, wherein the input of the noise estimation model is a signal and the output is a noise estimation value of the signal;

[0046] A training module, using the noise injection signal to train the noise estimation model;

[0047] The signal generation module selects a pure noise signal and inputs the pure noise signal into a trained noise estimation model to obtain a noise estimation value of the pure noise signal; and performs denoising processing on the pure noise signal according to the noise estimation value to obtain a generated signal similar to the target signal.

[0048] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0049] 1. In the present invention, the management of target signals without feedback information or prior knowledge is achieved based on the diffusion model. Specifically, noise is first added to the target signal based on the diffusion process, and then the signal with added noise is input into the constructed noise estimation model for noise estimation to achieve feature learning of the target signal. Then, the pure noise signal is denoised based on the learned features based on the inverse process. In this way, a signal similar to the target signal can be generated without prior knowledge or feedback information, thereby achieving management of dangerous signals.

[0050] 2. The noise estimation model of the present invention adopts TUnet network, which is an improved Unet network, realizes automatic feature extraction and optimization, can effectively separate noise and target signal, obtain accurate noise estimation value, and is conducive to high-quality generation of new signals.

[0051] Specifically, the TUnet network of the present invention is built with a residual network as a basic module, so that the network can learn the deep features of the signal, and solve the gradient disappearance problem in the model at the same time. The downsampling module helps to capture the contextual information of the signal and reduce the computational complexity. The intermediate processing network block helps to process and convert the features in depth, and the upsampling module helps to restore the detailed information of the signal, thereby achieving accurate positioning and recognition of the signal. The output network block further extracts and processes the upsampled data to generate the final output signal. The time information embedding network block embeds time information in the signal processing process. By incorporating time information into the network, the sensitivity of the model to dynamic changes in time can be improved, thereby obtaining better performance in time series signal analysis tasks. Through the organic combination of these network blocks, the TUnet network achieves efficient processing and feature extraction of input signals, and improves the overall performance and adaptability of the TUnet network.

[0052] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0054] Figure 1 It is a flow chart of a method for intelligently generating signals based on a diffusion model according to an embodiment of the present invention;

[0055] Figure 2 A schematic diagram of the structure of a TUnet network according to an embodiment of the present invention;

[0056] Figure 3 is a schematic diagram of a diffusion process according to an embodiment of the present invention;

[0057] Figure 4 is a schematic diagram of a reverse process of an embodiment of the present invention;

[0058] Figure 5 A curve diagram showing the change of noise loss value during model training according to an embodiment of the present invention;

[0059] Figure 6 It is a KDE probability distribution diagram of the preprocessed signal and the generated signal of an embodiment of the present invention;

[0060] Figure 7 It is a KL divergence value curve diagram of the preprocessed signal and the generated signal according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0062] An embodiment of the present invention discloses a method for intelligently generating signals based on a diffusion model. Figure 1 The method comprises:

[0063] Step 1, preprocessing the target signal to obtain a preprocessed signal;

[0064] Step 2, adding noise to the preprocessed signal based on a diffusion process to obtain a noise injection signal;

[0065] Step 3, constructing a noise estimation model, and using the noise injection signal to train the noise estimation model;

[0066] Step 4, selecting a pure noise signal, and inputting the pure noise signal into a trained noise estimation model to obtain a noise estimation value of the pure noise signal;

[0067] Step 5: Based on the inverse process, denoising is performed on the pure noise signal according to the noise estimation value, so as to obtain a generated signal similar to the target signal.

[0068] In the present invention, a similar signal of a target signal is generated based on a diffusion model. The core idea of ​​the diffusion model is to simulate the diffusion and reverse diffusion process of a physical process to generate data. In other words, the diffusion model includes two processes: a diffusion process (also called a forward process) and a reverse process (also called a reverse process).

[0069] In an embodiment of the present invention, a diffusion process is first simulated to add noise to a target signal to form a noise injection signal, and then the injected noise signal is used to train a noise estimation model to learn the characteristics of the target signal (mainly the probability distribution characteristics), and then the trained noise estimation model is used to obtain a noise estimation value in a pure noise signal, and then a reverse process is simulated to obtain a generated signal similar to the target signal based on the pure noise signal.

[0070] Compared with the prior art, in an embodiment of the present invention, management of target signals without feedback information or prior knowledge is achieved based on a diffusion model. Specifically, noise is first added to the target signal based on a diffusion process, and then the signal with added noise is input into a constructed noise estimation model for noise estimation to achieve feature learning of the target signal. Then, based on the inverse process, the pure noise signal is denoised using the learned features. In this way, a signal similar to the target signal can be generated without prior knowledge or feedback information, thereby achieving management of dangerous signals.

[0071] Step 1: preprocess the target signal to obtain a preprocessed signal.

[0072] The method further includes: collecting wireless signal data to be analyzed, and selecting a signal with a preset signal-to-noise ratio in a preset modulation mode from the wireless signal data as a target signal.

[0073] In the embodiment of the present invention, considering that the accuracy of signal processing is highly correlated with the signal-to-noise ratio, a signal with a preset signal-to-noise ratio in a preset modulation mode is selected as the target signal.

[0074] During implementation, a BPSK modulation pattern signal-to-noise ratio of 18 dB is selected as the target signal.

[0075] Specifically, the target signal is preprocessed, including:

[0076] The wireless signal data is normalized, wherein the normalization operation interval is [-1, 1].

[0077] The target signal is expressed as , the preprocessed signal is expressed as .

[0078] Step 2: adding noise to the preprocessed signal based on a diffusion process to obtain a noise injection signal.

[0079] Specifically, the diffusion process is a fixed Markov chain propagation process. Based on the fixed Markov chain propagation process, Iterate over the preprocessed signal Add noise to , thereby obtaining the noise injection signal , the formula is:

[0080] ; ;

[0081] ;

[0082] ;

[0083] In the formula, is an integer greater than 1, , The preprocessed signals are After t times, The signal after adding noise in iterations is is the iteration parameter, express It conforms to the standard normal distribution, and Draw random values ​​from a standard normal distribution, is the variance of the standard normal distribution at the t-th iteration, Randomly select a value from the variance value set, that is, .

[0084] In the embodiment of the present invention, when preprocessing the signal The Gaussian noise gradually meets the standard through T iterations until the cumulative noise value reaches , forming an injected noise signal ,like Figure 3 as shown in .

[0085] Among them, the fixed Markov chain propagation process usually refers to a random process with Markov properties, which can describe the transmission and change of signals. By simulating the fixed Markov chain propagation process, it is helpful to better analyze the signal.

[0086] Step 3: construct a noise estimation model, and use the noise injection signal to train the noise estimation model.

[0087] Wherein, by training the noise estimation model, the noise estimation model can accurately predict the noise added to the target signal.

[0088] Specifically, the loss function of the noise estimation model is:

[0089] ;

[0090] ;

[0091] In the formula, is the noise loss, For signal The noise estimation value obtained by inputting the noise estimation model, Represents the L2 loss function operation.

[0092] In the embodiment of the present invention, the loss function of the noise estimation model can accurately express the deviation between the real noise in the input signal and the noise estimation value output by the model.

[0093] Figure 5 FIG. 2 shows the change of the noise loss value during model training of an embodiment of the present invention.

[0094] In a specific embodiment, the noise estimation model is a TUnet network.

[0095] The TUnet network includes:

[0096] Residual network, used for preliminary feature extraction of input signals;

[0097] A downsampling network block, used to perform a downsampling operation on the data output by the residual network;

[0098] An intermediate processing network block, used for performing a deconvolution operation on the data output by the downsampling network block;

[0099] An upsampling network block, used for performing an upsampling operation on the data output by the intermediate processing network block and the data output by the downsampling network block;

[0100] A time information embedding network block is used to input time information into the upsampling network block, thereby embedding time information in output data of the upsampling network block and obtaining upsampling data embedded with time information;

[0101] The output network block is used to further extract features from the upsampled data embedded with the time information to generate a final output signal.

[0102] The noise estimation model of the embodiment of the present invention adopts the TUnet network, which is an improved Unet network, realizes automatic feature extraction and optimization, can effectively separate noise and target signals, obtain accurate noise estimation values, and is conducive to achieving high-quality generation of new signals.

[0103] Specifically, the TUnet network of the present invention is built with a residual network as a basic module, so that the network can learn the deep features of the signal, and solve the gradient disappearance problem in the model. The downsampling module helps to capture the contextual information of the signal and reduce the computational complexity. The intermediate processing network block helps to process and convert the features in depth. The upsampling module helps to restore the detailed information of the signal, thereby achieving accurate positioning and recognition of the signal. The output network block further extracts and processes the upsampled data to generate the final output signal. The time information embedding network block embeds time information in the signal processing process. By incorporating time information into the network, the sensitivity of the model to dynamic changes in time can be improved, thereby obtaining better performance in time series signal analysis tasks. Through the organic combination of these network blocks, the TUnet network achieves efficient processing and feature extraction of input signals, and improves the overall performance and adaptability of the TUnet network.

[0104] To be more specific, the residual network init_conv includes a 6-layer network connected in sequence, Lay0 to Lay5, where Lay0 is a one-dimensional convolution layer; Lay1 is a one-dimensional batch normalization layer, Lay2 is an activation layer, using a GELU activation function, Lay3 is a one-dimensional convolution layer, Lay4 is a one-dimensional batch normalization layer, and Lay5 is an activation layer, using a GELU activation function.

[0105] The downsampling network block Down includes two sequentially connected downsampling sub-network blocks, namely Down1 and Down2. Each downsampling sub-network block includes a 7-layer network, Lay0 to Lay6, which are sequentially connected. In each downsampling sub-network block, the Lay0 layer is a one-dimensional convolution layer, the Lay1 layer is a one-dimensional batch normalization layer, the Lay2 layer is an activation layer using the GELU activation function, the Lay3 layer is a one-dimensional convolution layer, the Lay4 layer is a one-dimensional batch normalization layer, the Lay5 layer is an activation layer using the GELU activation function, and the Lay6 layer is a maximum pooling layer.

[0106] The intermediate processing network block includes 5 layers of networks connected in sequence, from Lay0 to Lay4. Lay0 is a one-dimensional mean downsampling layer, Lay1 is an activation layer that uses the GELU activation function, Lay2 is a one-dimensional deconvolution network layer, Lay3 is a stack normalization layer, and Lay4 is an activation layer that uses the GELU activation function.

[0107] The upsampling network block Up includes two upsampling sub-network blocks connected in sequence, and each upsampling sub-network block includes 13 layers of networks connected in sequence, from Lay0 to Lay12. In each upsampling sub-network block, Lay0 is a one-dimensional inverse convolutional network layer, Lay1 is a one-dimensional convolutional layer, Lay2 is a one-dimensional batch normalization layer, Lay3 is an activation function layer, Lay4 is a one-dimensional convolutional layer, Lay5 is a one-dimensional batch normalization layer, Lay6 is an activation layer using a GELU activation function, Lay7 is a one-dimensional convolutional layer, Lay8 is a one-dimensional batch normalization layer, Lay9 is an activation layer using a GELU activation function, Lay10 is a one-dimensional convolutional layer, Lay11 is a one-dimensional batch normalization layer, and Lay12 is an activation layer using a GELU activation function.

[0108] The output network block Out includes four layers of networks connected in sequence, from Lay0 to Lay3, where Lay0 is a one-dimensional convolution layer, Lay1 is a stack normalization layer, Lay2 is an activation layer using a GELU activation function layer, and Lay3 is a one-dimensional convolution layer. The output network block outputs a signal of the same size as the input signal.

[0109] The time information embedding network block Timeembed includes two time information embedding sub-network blocks, each of which has three layers of networks. Layer Lay0 is a one-dimensional linear transformation network layer, Layer Lay1 is an activation layer that uses the GELU activation function, and Layer Lay2 is a one-dimensional linear transformation network layer.

[0110] like Figure 2 As shown in , the output of the residual network init_conv is connected to the input of the first downsampling sub-network block Down1, the output of the first downsampling sub-network block Down1 is connected to the second downsampling sub-network block Down2, the output of the first downsampling sub-network block Down1 is also connected to the second upsampling sub-network block Up2, the output of the second downsampling sub-network block Down2 is connected to the intermediate processing network block, the output of the second downsampling sub-network block Down2 is also connected to the first upsampling sub-network block Up1, the output of the intermediate processing network block is connected to the first upsampling sub-network block Up1, and the output of the second upsampling sub-network block Up2 is connected to the output network block Out. At the same time, the first time information embedding sub-network block Timeembed1 is connected to the first sampling sub-network block Up1, so that the time information is embedded in the output data of the first sampling sub-network block Up1; the second time information embedding sub-network block Timeembed2 is connected to the second sampling sub-network block Up2, so that the time information is embedded in the output data of the first sampling sub-network block Up2.

[0111] Step 4: Select a pure noise signal, and input the pure noise signal into a trained noise estimation model to obtain a noise estimation value of the pure noise signal.

[0112] Specifically, the pure noise signal is a random noise signal, and a signal value of the random noise signal is randomly taken from a standard normal distribution.

[0113] Among them, the trained noise estimation model performs deep learning on the characteristics of the target signal, and obtains a noise estimation value by inputting a pure noise signal into the trained noise estimation model, which can be used to generate a signal with the same or similar probability distribution as the target signal.

[0114] Step 5: Based on the inverse process, denoising is performed on the pure noise signal according to the noise estimation value, so as to obtain a generated signal similar to the target signal.

[0115] Specifically, the reverse process is a learnable Markov chain propagation process.

[0116] Wherein, based on the learnable Markov chain propagation process, the pure noise signal is Perform T steps of iteration to perform denoising processing, thereby obtaining a generated signal similar to the target signal , the formula is:

[0117] ; ;

[0118] ;

[0119] In the formula, , Pure noise signal go through , The signal after the denoising process is performed iteratively. for Input the signal noise estimate obtained from the trained noise estimation model, yes The cumulative mean of the standard normal distribution of iterations; express It conforms to the standard normal distribution, and hour, Draw random values ​​from a standard normal distribution; hour, .

[0120] In the embodiment of the present invention, the reverse process is simulated, and the noise removal operation is gradually performed through T iterations to make it closer to the original pre-processed signal. , until the noise estimation value obtained by using the trained noise estimation model is removed, thereby obtaining a generated signal similar to the target signal, such as Figure 4 as shown in .

[0121] In one embodiment, the method further comprises:

[0122] Kernel density estimation is performed on the generated signal and the preprocessed signal respectively, and corresponding probability density curves are drawn, such as Figure 6 As shown in;

[0123] The probability density curves of the generated signal and the preprocessed signal are compared to determine whether the probability distributions of the two signals are the same, so as to evaluate the similarity between the generated signal and the target signal.

[0124] Kernel density estimation (KDE) can be used to compare whether the probability distributions of two signals are the same, thereby evaluating their similarity. If the probability density curves of two signals are similar, then the two signals can be considered statistically similar.

[0125] The basic principle of kernel density estimation is to place a kernel function around each data point, then add up all the kernel functions and normalize them to get a smooth probability density estimate.

[0126] For a one-dimensional time series signal , the calculation formula of kernel density estimation is:

[0127] ;

[0128] ;

[0129] ;

[0130] in, is the Gaussian kernel function, Yes The estimated density at .

[0131] In one embodiment, the method further comprises:

[0132] The KL divergence between the generated signal and the preprocessed signal is calculated to evaluate the similarity between the generated signal and the target signal.

[0133] KL divergence (Kullback-Leibler Divergence), also known as relative entropy, is used to measure the difference between two probability distributions P and Q.

[0134] Figure 7 Graph 1 shows the KL divergence value curves of the preprocessed signal and the generated signal of an embodiment of the present invention.

[0135] The calculation formula of KL divergence is:

[0136] ;

[0137] In the formula, and are the discrete probability distribution functions of the generated signal and the target signal respectively. When the probability distributions of the two signals are the same, the KL divergence is 0.

[0138] On the other hand, an embodiment of the present invention further provides a signal intelligent generation device based on a diffusion model. The device comprises:

[0139] A preprocessing module preprocesses the target signal without prior knowledge to obtain a preprocessed signal;

[0140] A noise injection module, adding noise to the preprocessed signal based on a diffusion process to obtain a noise injection signal;

[0141] A noise estimation model, wherein the input of the noise estimation model is a signal and the output is a noise estimation value of the signal;

[0142] A training module, using the noise injection signal to train the noise estimation model;

[0143] The signal generation module selects a pure noise signal and inputs the pure noise signal into a trained noise estimation model to obtain a noise estimation value of the pure noise signal; based on a reverse process, denoising is performed on the pure noise signal according to the noise estimation value to obtain a generated signal similar to the target signal.

[0144] The signal intelligent generation device based on diffusion model of the present invention is used to implement the signal intelligent generation method based on diffusion model. The device and method of the present invention have the same technical features and can achieve the same technical effects, which will not be described in detail here.

[0145] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0146] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A signal intelligent generation method based on a diffusion model, characterized in that: The method comprises: Preprocessing the target signal to obtain a preprocessed signal; Adding noise to the preprocessed signal based on a diffusion process to obtain a noise injection signal; Constructing a noise estimation model, and using the noise injection signal to train the noise estimation model; Selecting a pure noise signal, and inputting the pure noise signal into a trained noise estimation model to obtain a noise estimation value of the pure noise signal; Based on the inverse process, denoising the pure noise signal according to the noise estimation value, so as to obtain a generated signal; The loss function of the noise estimation model is: ； ； In the formula, is the noise loss, For signal The noise estimation value obtained by inputting the noise estimation model, Represents the L2 loss function operation; The noise estimation model is a TUnet network, and the TUnet network includes: Residual network, used for preliminary feature extraction of input signals; A downsampling network block, used to perform a downsampling operation on the data output by the residual network; An intermediate processing network block, used for performing a deconvolution operation on the data output by the downsampling network block; An upsampling network block, used for performing an upsampling operation on the data output by the intermediate processing network block and the data output by the downsampling network block; A time information embedding network block is used to input time information into the upsampling network block, thereby embedding time information in output data of the upsampling network block and obtaining upsampling data embedded with time information; The output network block is used to further extract features from the upsampled data embedded with the time information to generate a final output signal.

2. The method for intelligently generating signals based on a diffusion model according to claim 1, characterized in that: The diffusion process is a fixed Markov chain propagation process; Among them, based on the fixed Markov chain propagation process, through Iterate over the preprocessed signal Add noise to , thereby obtaining the noise injection signal , the formula is: ； ； ； ； In the formula, is an integer greater than 1, , The preprocessed signals are After t times, The signal after adding noise iteratively is is the iteration parameter, express It conforms to the standard normal distribution, and Draw random values ​​from a standard normal distribution, is the variance of the standard normal distribution at the t-th iteration, Pick a random value from the set of variance values.

3. The method for intelligently generating signals based on a diffusion model according to claim 2, characterized in that: The reverse process is a learnable Markov chain propagation process; Wherein, based on the learnable Markov chain propagation process, the pure noise signal is Perform T steps of iterations to perform denoising, thereby obtaining the generated signal , the formula is: ； ； ； In the formula, , Pure noise signal go through , The signal after the denoising process is performed iteratively. for Input the signal noise estimate obtained from the trained noise estimation model, yes The cumulative mean of the standard normal distribution of iterations; express It conforms to the standard normal distribution, and hour, Draw random values ​​from a standard normal distribution; hour, .

4. A method for intelligently generating signals based on a diffusion model according to any one of claims 1 to 3, characterized in that: The pure noise signal is a random noise signal, and the signal value of the random noise signal is randomly taken from a standard normal distribution.

5. A method for intelligently generating signals based on a diffusion model according to any one of claims 1 to 3, characterized in that: The method comprises: Wireless signal data to be analyzed is collected, and a signal with a preset signal-to-noise ratio in a preset modulation mode is selected from the wireless signal data as a target signal.

6. A method for intelligently generating signals based on a diffusion model according to any one of claims 1 to 3, characterized in that: The method further comprises: Performing kernel density estimation on the generated signal and the preprocessed signal respectively, and drawing corresponding probability density curves; The probability density curves of the generated signal and the preprocessed signal are compared to determine whether the probability distributions of the two signals are the same, so as to evaluate the similarity between the generated signal and the target signal.

7. A method for intelligently generating signals based on a diffusion model according to any one of claims 1 to 3, characterized in that: The method further comprises: The KL divergence between the generated signal and the preprocessed signal is calculated to evaluate the similarity between the generated signal and the target signal.

8. A signal intelligent generation device based on a diffusion model, characterized in that: The device comprises: A preprocessing module preprocesses the target signal without prior knowledge to obtain a preprocessed signal; A noise injection module, adding noise to the preprocessed signal based on a diffusion process to obtain a noise injection signal; A noise estimation model, wherein the input of the noise estimation model is a signal and the output is a noise estimation value of the signal; A training module, using the noise injection signal to train the noise estimation model; A signal generation module selects a pure noise signal and inputs the pure noise signal into a trained noise estimation model to obtain a noise estimation value of the pure noise signal; based on a reverse process, denoises the pure noise signal according to the noise estimation value to obtain a generated signal; The loss function of the noise estimation model is: ； ； In the formula, is the noise loss, For signal The noise estimation value obtained by inputting the noise estimation model, Represents the L2 loss function operation; The noise estimation model is a TUnet network, and the TUnet network includes: Residual network, used for preliminary feature extraction of input signals; A downsampling network block, used to perform a downsampling operation on the data output by the residual network; An intermediate processing network block, used for performing a deconvolution operation on the data output by the downsampling network block; An upsampling network block, used for performing an upsampling operation on the data output by the intermediate processing network block and the data output by the downsampling network block; A time information embedding network block is used to input time information into the upsampling network block, thereby embedding time information in output data of the upsampling network block and obtaining upsampling data embedded with time information; The output network block is used to further extract features from the upsampled data embedded with the time information to generate a final output signal.

Images