Modulation signal generation method based on diffusion model

Abstract

The invention discloses a modulation signal generation method based on a diffusion model, and relates to the modulation signal generation method based on the diffusion model. The invention aims to solve the problems that when sample data of various modulation signals are actually collected, the problems such as signal time synchronization, surrounding environment change and equipment debugging need to be solved with a large amount of cost, and signal collection and storage involve a large amount of learning cost and time cost. The method comprises the following steps: 1, acquiring paired simulation signals and real signals; 2, obtaining a simulation signal and a real signal after standardization processing; 3, acquiring a data set for a PSK (Phase Shift Keying) group modulation mode, a data set for a QAM (Quadrature Amplitude Modulation) group modulation mode and a data set for an APSK (Amplitude Phase Shift And 4, based on the standardized simulation signals, aiming at the data sets of the PSK group modulation mode, the QAM group modulation mode and the APSK group modulation mode, adopting a diffusion model to generate different modulation signals. The method is applied to the field of modulation signal generation.

Description

Technical Field

[0001] The present invention relates to a method for generating a modulated signal based on a diffusion model. Background Art

[0002] The diffusion model is a type of generative model, and its core is to generate high-quality data samples by gradually denoising. The main steps are divided into two steps: the forward diffusion process, which is defined as a series of conditional distributions q(x t |x t-1 ), starting from the real data distribution and gradually transforming towards the noise distribution. By gradually adding noise, the data samples are transformed into pure noise; the reverse denoising process, which is also defined as a series of conditional distributions p θ (x t-1 |x t ), and the parameter θ is learned by a neural network. Starting from pure noise, the model reversely predicts and removes the noise step by step according to the learned parameters, and restores to the original data distribution, thereby generating high-quality samples. The overall process of the model is as Figure 2 shown. Due to its generation quality and flexibility, it has become another powerful generative model besides generative adversarial networks (GANs).

[0003] In recent years, diffusion models have been widely applied to tasks such as generating natural language texts and image generation tasks, and have achieved good results. However, at present, there is a lack of a solution for generating I / Q signal data using diffusion models; in addition, when actually collecting sample data of various modulated signals, due to the complexity of the actual environment and the reasons of the hardware device itself, it is necessary to spend a large amount of cost to solve problems such as signal time synchronization, changes in the surrounding environment, and device debugging. The acquisition and storage of signals of each modulation method will involve a large amount of learning cost and time cost. Summary of the Invention

[0004] The purpose of the present invention is to solve the problem that when actually collecting sample data of various modulated signals, it is necessary to spend a large amount of cost to solve problems such as signal time synchronization, changes in the surrounding environment, and device debugging. The acquisition and storage of signals of each modulation method will involve a large amount of learning cost and time cost, and to propose a method for generating a modulated signal based on a diffusion model.

[0005] The specific process of a method for generating a modulated signal based on a diffusion model is as follows:

[0006] Step 1: Obtain paired simulation signals and real signals;

[0007] Step 2: Standardize the simulated signal and the real signal obtained in Step 1 to obtain the standardized simulated signal and real signal;

[0008] Step 3: Based on the real signal after standardization in Step 2, obtain datasets for PSK modulation schemes, QAM modulation schemes, and APSK modulation schemes;

[0009] Step 4: Based on the standardized simulated signal in Step 2 and the datasets for PSK modulation schemes, QAM modulation schemes, and APSK modulation schemes obtained in Step 3, use a diffusion model to generate different modulation signals; the specific process is as follows:

[0010] Step 4-1:

[0011] Perform forward diffusion on the signal data in the dataset of the PSK modulation scheme to obtain the signal data after forward diffusion, and save the signal data x t , time step t, and the actually added noise ε;

[0012] Perform forward diffusion on the signal data in the dataset of the QAM modulation scheme to obtain the signal data after forward diffusion, and save the signal data x t , time step t, and the actually added noise ε;

[0013] Perform forward diffusion on the signal data in the dataset of the APSK modulation scheme to obtain the signal data after forward diffusion, and save the signal data x t , time step t, and the actually added noise ε;

[0014] Step 4-2:

[0015] Take the signal data x t after forward diffusion obtained from the PSK modulation scheme, time step t, and the standardized simulated signal in Step 2 as the input of the U-Net network, and take the predicted noise ε θ added at time step t as the output of the U-Net network. Based on the real noise ε saved during the forward diffusion process and the predicted noise ε θ added at time step t, calculate the loss function L until the loss function L converges to obtain the trained U-Net network for the PSK modulation scheme;

[0016] Take the signal data x t after forward diffusion obtained from the QAM modulation scheme, time step t, and the standardized simulated signal in Step 2 as the input of the U-Net network, and take the predicted noise ε θAs the output of the U-Net network, based on the true noise ε saved during the forward diffusion process and the noise ε added at the predicted time step t θ Calculate the loss function L until the loss function L converges to obtain a trained U-Net network for the QAM group modulation method;

[0017] The signal data x after forward diffusion obtained by the APSK group modulation method t , the time step t, and the simulated signal after the normalization process in step two are used as the input of the U-Net network, and the noise ε added at the predicted time step t θ As the output of the U-Net network, based on the true noise ε saved during the forward diffusion process and the noise ε added at the predicted time step t θ Calculate the loss function L until the loss function L converges to obtain a trained U-Net network for the APSK group modulation method;

[0018] Step Four Three.

[0019] Based on the trained U-Net network for the PSK modulation method, the normalized 32PSK simulated signal is input into the trained U-Net network for the PSK modulation method, and starting from the pure noise x T Perform reverse denoising, passing through x T , x T-1 , x T-2 , …, x1, to generate the predicted value of the final 32PSK true signal; perform denormalization on the predicted value of the finally generated 32PSK true signal data to obtain the predicted value of the finally generated 32PSK true signal data after denormalization;

[0020] Based on the trained U-Net network for the QAM modulation method, the normalized 256QAM simulated signal is input into the trained U-Net network for the QAM modulation method, and starting from the pure noise x T Perform reverse denoising, passing through x T , x T-1 , x T-2 , …, x1, to generate the predicted value of the final 256QAM true signal; perform denormalization on the predicted value of the finally generated 256QAM true signal data to obtain the predicted value of the finally generated 256QAM true signal data after denormalization;

[0021] Based on the trained U-Net network for the APSK modulation method, the normalized 128APSK simulated signal is input into the trained U-Net network for the APSK modulation method, and starting from the pure noise x T Perform reverse denoising, passing through x T , x T-1 , xT-2 …, x1 to generate the predicted value of the final 128 - APSK true signal; perform inverse normalization on the predicted value of the finally generated 128 - APSK true signal data to obtain the predicted value of the finally generated 128 - APSK true signal data after inverse normalization.

[0022] The beneficial effects of the present invention are as follows:

[0023] The main content involved in the present invention is a method for generating modulation signals based on a diffusion model. The purpose is to generalize and generate some specific types of modulation signals through the diffusion model, so as to perform signal modulation recognition on the generated signals and existing signals together.

[0024] For signal modulation recognition based on a generative model, the main process is: determining the modulation method for training the diffusion model and the modulation method to be generated by the diffusion model; using some simulation signals and corresponding true signals passing through a specific environment to train the diffusion model; using the trained diffusion model to generate the generated signals of the required modulation method; saving the signals in the same form as the training signals to complete the generation and saving of specific modulation signals.

[0025] The present invention proposes a method for generating modulation signals based on a diffusion model. It is intended to use the diffusion model to learn the mapping relationship of a specific channel environment through existing paired simulation signals and true signals. Thus, when we need other similar types of true modulation signals again, we can directly use the trained diffusion model for generalization generation, eliminating the need to redeploy and collect through hardware devices, saving a large amount of time and improving efficiency. In this way, the cost of collecting true signals is greatly saved, which is of great significance for the generation and expansion of the modulation signal dataset. Brief Description of the Drawings

[0026] Figure 1 is the flowchart of the present invention;

[0027] Figure 2 is the overall flowchart of the diffusion model;

[0028] Figure 3 is the diagram of the simulation signal and true signal of 32 - APSK. a is the simulation signal of 32 - APSK, and b is the true signal of 32 - APSK;

[0029] Figure 4 is the structure diagram of the U - Net network;

[0030] Figure 5 is the overall flowchart of the reverse denoising process;

[0031] Figure 6 It is a confusion matrix diagram. Detailed implementation manners

[0032] Detailed implementation manner 1: The specific process of a modulation signal generation method based on a diffusion model in this implementation manner is as follows:

[0033] Step 1: Obtain paired simulation signals and real signals;

[0034] Step 2: Perform normalization processing on the simulation signals and real signals obtained in Step 1 to obtain the normalized simulation signals and real signals;

[0035] Step 3: Based on the real signals after normalization processing in Step 2, obtain data sets for PSK group modulation mode, QAM group modulation mode, and APSK group modulation mode;

[0036] Step 4: Based on the simulation signals after normalization processing in Step 2 and the data sets for PSK group modulation mode, QAM group modulation mode, and APSK group modulation mode obtained in Step 3, use a diffusion model to generate different modulation signals; the specific process is as follows:

[0037] Step 4-1:

[0038] Perform forward diffusion on the signal data in the data set of the PSK group modulation mode to obtain the signal data after forward diffusion, and save the signal data x t , time step t, and the actually added noise ε;

[0039] Perform forward diffusion on the signal data in the data set of the QAM group modulation mode to obtain the signal data after forward diffusion, and save the signal data x t , time step t, and the actually added noise ε;

[0040] Perform forward diffusion on the signal data in the data set of the APSK group modulation mode to obtain the signal data after forward diffusion, and save the signal data x t , time step t, and the actually added noise ε;

[0041] Step 4-2:

[0042] Use the signal data x t after forward diffusion obtained in the PSK group modulation mode, time step t, and the simulation signals after normalization processing in Step 2 as the input of the U-Net network, and use the predicted noise ε θ added at time step t as the output of the U-Net network. Based on the real noise ε saved during the forward diffusion process and the predicted noise ε θCalculate the loss function L until the loss function L converges, and obtain a trained U-Net network for the PSK group modulation method;

[0043] Take the signal data x after forward diffusion obtained by the QAM group modulation method t , the time step t, and the simulated signal after the normalization process in step two as the input of the U-Net network, and take the predicted noise ε added at the t time step θ as the output of the U-Net network. Based on the true noise ε saved during the forward diffusion process and the predicted noise ε added at the t time step θ calculate the loss function L until the loss function L converges, and obtain a trained U-Net network for the QAM group modulation method;

[0044] Take the signal data x after forward diffusion obtained by the APSK group modulation method t , the time step t, and the simulated signal after the normalization process in step two as the input of the U-Net network, and take the predicted noise ε added at the t time step θ as the output of the U-Net network. Based on the true noise ε saved during the forward diffusion process and the predicted noise ε added at the t time step θ calculate the loss function L until the loss function L converges, and obtain a trained U-Net network for the APSK group modulation method;

[0045] Step Four Three.

[0046] Based on the trained U-Net network for the PSK modulation method, the normalized 32PSK simulated signal is input into the trained U-Net network for the PSK modulation method, and start reverse denoising from the pure noise x T , through x T , x T-1 , x T-2 , …, x1, to generate the predicted value of the final 32PSK true signal; perform inverse normalization (the inverse normalization in step 2) on the predicted value of the final generated 32PSK true signal data to obtain the predicted value of the final generated 32PSK true signal data after inverse normalization;

[0047] Based on the trained U-Net network for the QAM modulation method, the normalized 256QAM simulated signal is input into the trained U-Net network for the QAM modulation method, and start reverse denoising from the pure noise x T , through x T , x T-1 , x T-2, …, x1, to generate the predicted value of the final 256QAM true signal; perform denormalization on the predicted value of the finally generated 256QAM true signal data (the denormalization in step 2) to obtain the predicted value of the finally generated 256QAM true signal data after denormalization;

[0048] Based on the trained U-Net network with APSK modulation, the normalized 128APSK simulation signal is input into the trained U-Net network with APSK modulation, starting from pure noise x T to start reverse denoising, passing through x T , x T-1 , x T-2 , …, x1, to generate the predicted value of the final 128APSK true signal; perform denormalization on the predicted value of the finally generated 128APSK true signal data (the denormalization in step 2) to obtain the predicted value of the finally generated 128APSK true signal data after denormalization.

[0049] Specific Embodiment 2: The difference between this embodiment and Specific Embodiment 1 is that in step 1, paired simulation signals and true signals are obtained; the specific process is as follows:

[0050] Step 1.1: Obtain simulation signals under different modulation methods;

[0051] Step 1.2: Obtain true signals corresponding to the simulation signals under different modulation methods.

[0052] Other steps and parameters are the same as those in Specific Embodiment 1.

[0053] Specific Embodiment 3: The difference between this embodiment and Specific Embodiment 1 or 2 is that in step 1.1, simulation signals under different modulation methods are obtained; the specific process is as follows:

[0054] Introduction of simulation signals

[0055] The forms of the simulation signals and true signals used in the present invention are both OFDM signals. The number of subcarriers used for each modulation method is 256, and 100,000 OFDM symbols are generated for each modulation method. The specific steps for generating the simulation signals are as follows:

[0056] Step 1.1.1:

[0057] Under the premise that the probability of each bit is 0.5;

[0058] Generate a sufficient number of random bitstreams for the QPSK modulation method. The bitstream consists of bits, which are data of 1 or 0. The probabilities of 0 and 1 appearing in the bitstream are the same (a bitstream composed of a bunch of 0 and 1 data). The sufficient number is to ensure that 100,000 OFDM symbols can be generated later;

[0059] Generate a sufficient number of random bit streams for BPSK modulation. The bit stream consists of bits, which are data of 1 or 0. The occurrence probabilities of 0 and 1 in the bit stream are the same (a bit stream is composed of a pile of 0 and 1 data). The sufficient number is to ensure that 100,000 OFDM symbols are generated later;

[0060] Generate a sufficient number of random bit streams for 8PSK modulation. The bit stream consists of bits, which are data of 1 or 0. The occurrence probabilities of 0 and 1 in the bit stream are the same (a bit stream is composed of a pile of 0 and 1 data). The sufficient number is to ensure that 100,000 OFDM symbols are generated later;

[0061] Generate a sufficient number of random bit streams for 16PSK modulation. The bit stream consists of bits, which are data of 1 or 0. The occurrence probabilities of 0 and 1 in the bit stream are the same (a bit stream is composed of a pile of 0 and 1 data). The sufficient number is to ensure that 100,000 OFDM symbols are generated later;

[0062] Generate a sufficient number of random bit streams for 32PSK modulation. The bit stream consists of bits, which are data of 1 or 0. The occurrence probabilities of 0 and 1 in the bit stream are the same (a bit stream is composed of a pile of 0 and 1 data). The sufficient number is to ensure that 100,000 OFDM symbols are generated later;

[0063] Generate a sufficient number of random bit streams for 16QAM modulation. The bit stream consists of bits, which are data of 1 or 0. The occurrence probabilities of 0 and 1 in the bit stream are the same (a bit stream is composed of a pile of 0 and 1 data). The sufficient number is to ensure that 100,000 OFDM symbols are generated later;

[0064] Generate a sufficient number of random bit streams for 64QAM modulation. The bit stream consists of bits, which are data of 1 or 0. The occurrence probabilities of 0 and 1 in the bit stream are the same (a bit stream is composed of a pile of 0 and 1 data). The sufficient number is to ensure that 100,000 OFDM symbols are generated later;

[0065] Generate a sufficient number of random bit streams for 256QAM modulation. The bit stream consists of bits, which are data of 1 or 0. The occurrence probabilities of 0 and 1 in the bit stream are the same (a bit stream is composed of a pile of 0 and 1 data). The sufficient number is to ensure that 100,000 OFDM symbols are generated later;

[0066] Generate a sufficient number of random bit streams for 16APSK modulation. The bit stream consists of bits, which are data of 1 or 0. The occurrence probabilities of 0 and 1 in the bit stream are the same (a bit stream is composed of a pile of 0 and 1 data). The sufficient number is to ensure that 100,000 OFDM symbols are generated later;

[0067] Generate a sufficient number of random bit streams for 32APSK modulation. The bit stream consists of bits, which are data of 1 or 0. The occurrence probabilities of 0 and 1 in the bit stream are the same (a bit stream is composed of a bunch of 0s and 1s). The sufficient number is to ensure the generation of 100,000 OFDM symbols later;

[0068] Generate a sufficient number of random bit streams for 64APSK modulation. The bit stream consists of bits, which are data of 1 or 0. The occurrence probabilities of 0 and 1 in the bit stream are the same (a bit stream is composed of a bunch of 0s and 1s). The sufficient number is to ensure the generation of 100,000 OFDM symbols later;

[0069] Generate a sufficient number of random bit streams for 128APSK modulation. The bit stream consists of bits, which are data of 1 or 0. The occurrence probabilities of 0 and 1 in the bit stream are the same (a bit stream is composed of a bunch of 0s and 1s). The sufficient number is to ensure the generation of 100,000 OFDM symbols later;

[0070] Step One One Two

[0071] For each modulation method, there is a corresponding different mapping rule from bits to symbols;

[0072] For QPSK modulation, every 2 bits are mapped to a QPSK complex number. One OFDM symbol is generated from every 256 QPSK complex numbers, and 100,000 OFDM symbols need to be generated;

[0073] For BPSK modulation, every 1 bit is mapped to a BPSK complex number. One OFDM symbol is generated from every 256 BPSK complex numbers, and 100,000 OFDM symbols need to be generated;

[0074] For 8PSK modulation, every 3 bits are mapped to an 8PSK complex number. One OFDM symbol is generated from every 256 8PSK complex numbers, and 100,000 OFDM symbols need to be generated;

[0075] For 16PSK modulation, every 4 bits are mapped to a 16PSK complex number. One OFDM symbol is generated from every 256 16PSK complex numbers, and 100,000 OFDM symbols need to be generated;

[0076] For 32PSK modulation, every 5 bits are mapped to a 32PSK complex number. One OFDM symbol is generated from every 256 32PSK complex numbers, and 100,000 OFDM symbols need to be generated;

[0077] For 16QAM modulation, every 4 bits are mapped to a 16QAM complex number, and one OFDM symbol is generated from every 256 16QAM complex numbers. 100,000 OFDM symbols need to be generated;

[0078] For 64QAM modulation, every 6 bits are mapped to a 64QAM complex number, and one OFDM symbol is generated from every 256 64QAM complex numbers. 100,000 OFDM symbols need to be generated;

[0079] For 256QAM modulation, every 8 bits are mapped to a 256QAM complex number, and one OFDM symbol is generated from every 256 256QAM complex numbers. 100,000 OFDM symbols need to be generated;

[0080] For 16APSK modulation, every 4 bits are mapped to a 16APSK complex number, and one OFDM symbol is generated from every 256 16APSK complex numbers. 100,000 OFDM symbols need to be generated;

[0081] For 32APSK modulation, every 5 bits are mapped to a 32APSK complex number, and one OFDM symbol is generated from every 256 32APSK complex numbers. 100,000 OFDM symbols need to be generated;

[0082] For 64APSK modulation, every 6 bits are mapped to a 64APSK complex number, and one OFDM symbol is generated from every 256 64APSK complex numbers. 100,000 OFDM symbols need to be generated;

[0083] For 128APSK modulation, every 7 bits are mapped to a 128APSK complex number, and one OFDM symbol is generated from every 256 128APSK complex numbers. 100,000 OFDM symbols need to be generated;

[0084] Step 1-13

[0085] For QPSK modulation, the 100,000 OFDM symbols generated in Step 1-12 are concatenated and saved as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal of QPSK modulation;

[0086] For BPSK modulation, the 100,000 OFDM symbols generated in Step 1-12 are concatenated and saved as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal of BPSK modulation;

[0087] For the 8PSK modulation method, the 100,000 OFDM symbols generated in Step 112 are concatenated and saved as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal for the 8PSK modulation method;

[0088] For the 16PSK modulation method, the 100,000 OFDM symbols generated in Step 112 are concatenated and saved as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal for the 16PSK modulation method;

[0089] For the 32PSK modulation method, the 100,000 OFDM symbols generated in Step 112 are concatenated and saved as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal for the 32PSK modulation method;

[0090] For the 16QAM modulation method, the 100,000 OFDM symbols generated in Step 112 are concatenated and saved as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal for the 16QAM modulation method;

[0091] For the 64QAM modulation method, the 100,000 OFDM symbols generated in Step 112 are concatenated and saved as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal for the 64QAM modulation method;

[0092] For the 256QAM modulation method, the 100,000 OFDM symbols generated in Step 112 are concatenated and saved as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal for the 256QAM modulation method;

[0093] For the 16APSK modulation method, the 100,000 OFDM symbols generated in Step 112 are concatenated and saved as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal for the 16APSK modulation method;

[0094] For the 32APSK modulation method, the 100,000 OFDM symbols generated in Step 112 are concatenated and saved as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal for the 32APSK modulation method;

[0095] For the 64APSK modulation method, the 100,000 OFDM symbols generated in step 112 are concatenated and saved as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal of the 64APSK modulation method;

[0096] For the 128APSK modulation method, the 100,000 OFDM symbols generated in step 112 are concatenated and saved as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal of the 128APSK modulation method;

[0097] That is, a total of 12 clean signals (simulation signals) of QPSK, BPSK, 8PSK, 16PSK, 32PSK, 16QAM, 64QAM, 256QAM, 16APSK, 32APSK, 64APSK, and 128APSK are finally saved, and the shape of each is (100,000×256).

[0098] Other steps and parameters are the same as those in the first or second specific implementation manner.

[0099] Specific implementation manner four: The difference between this implementation manner and one of the first to third specific implementation manners is that in step 12, the real signal corresponding to the simulation signal under different modulation methods is obtained; the specific process is as follows:

[0100] After the generation and saving of the simulation signal are completed through the above steps, it is necessary to use each simulation signal of the 12 modulation signals saved to generate the corresponding real signal. The specific generation steps are as follows:

[0101] Step 121: For the QPSK modulation method, obtain the real signal corresponding to the simulation signal under the QPSK modulation method;

[0102] Step 122: For the BPSK modulation method, obtain the real signal corresponding to the simulation signal under the BPSK modulation method; the process is the same as step 121;

[0103] Step 123: For the 8PSK modulation method, obtain the real signal corresponding to the simulation signal under the 8PSK modulation method; the process is the same as step 121;

[0104] Step 124: For the 16PSK modulation method, obtain the real signal corresponding to the simulation signal under the 16PSK modulation method; the process is the same as step 121;

[0105] Step 125: For the 32PSK modulation method, obtain the real signal corresponding to the simulated signal under the 32PSK modulation method; the process is the same as that in Step 121;

[0106] Step 126: For the 16QAM modulation method, obtain the real signal corresponding to the simulated signal under the 16QAM modulation method; the process is the same as that in Step 121;

[0107] Step 127: For the 64QAM modulation method, obtain the real signal corresponding to the simulated signal under the 64QAM modulation method; the process is the same as that in Step 121;

[0108] Step 128: For the 256QAM modulation method, obtain the real signal corresponding to the simulated signal under the 256QAM modulation method; the process is the same as that in Step 121;

[0109] Step 129: For the 16APSK modulation method, obtain the real signal corresponding to the simulated signal under the 16APSK modulation method; the process is the same as that in Step 121;

[0110] Step 120: For the 32APSK modulation method, obtain the real signal corresponding to the simulated signal under the 32APSK modulation method; the process is the same as that in Step 121;

[0111] Step 121: For the 64APSK modulation method, obtain the real signal corresponding to the simulated signal under the 64APSK modulation method; the process is the same as that in Step 121;

[0112] Step 122: For the 128APSK modulation method, obtain the real signal corresponding to the simulated signal under the 128APSK modulation method; the process is the same as that in Step 121.

[0113] Other steps and parameters are the same as those in any one of Embodiments 1 to 3.

[0114] Specific Embodiment 5: The difference between this embodiment and any one of Embodiments 1 to 4 is that in Step 121 for the QPSK modulation method, the real signal corresponding to the simulated signal under the QPSK modulation method is obtained;

[0115] The specific process is as follows:

[0116] 1) Perform an inverse discrete Fourier transform IDFT on each OFDM symbol in the simulated signal generated by the QPSK modulation method, and add a cyclic prefix CP to each OFDM symbol after the inverse discrete Fourier transform IDFT to obtain each OFDM symbol with the cyclic prefix CP added, thus completing the work before OFDM enters the channel;

[0117] 2) Each OFDM symbol after adding the cyclic prefix CP passes through the channel to obtain the OFDM symbol affected by the channel;

[0118] The specific process is as follows:

[0119] 21) Let i = 1

[0120] 22) Convolve the i-th OFDM symbol after adding the cyclic prefix CP with the channel impulse response to obtain the convolved signal; the specific process is as follows:

[0121] The open-source Htrain.mat dataset contains 9000 groups of data. The shape of each group of data is 64×126×2, where 64 is the number of channels, 126 is the delay, and 2 is two-way IQ (complex numbers);

[0122] IQ modulation means that the data is divided into two paths and carrier modulation is performed separately. The two carriers are orthogonal to each other. I: in-phase, Q: quadrature.

[0123] IQ modulation is a problem of the direction of vectors. The in-phase is the signal with the same vector direction; the quadrature component is that the two signal vectors are orthogonal (differ by 90°); the IQ signal is that one path is 0° or 180°, and the other path is 90° or 270°, which are called the I path and the Q path, and they are two orthogonal signals.

[0124] Select the data of the first channel of each group of data in the 9000 groups of data as the channel data H. The shape of the channel data H is 9000×126×2, 126 is the delay, and 2 is two-way IQ (complex numbers);

[0125] Perform convolution calculation on the i-th OFDM symbol and a randomly selected group of data 126×2 in the 9000 groups;

[0126] This is to fully cover and affect the 9000 groups of channel data;

[0127] 23) Add Gaussian noise to the signal after convolution calculation at a signal-to-noise ratio of 70 dB to obtain the signal after adding Gaussian noise;

[0128] 24) Remove the cyclic prefix CP from the signal after adding Gaussian noise, and perform a fast Fourier transform DFT on the signal after removing the cyclic prefix CP to obtain the OFDM symbol affected by the channel (consisting of 256 complex numbers);

[0129] 25) Let \(i = i + 1\), and repeat steps 22) to 24) until all 100,000 OFDM symbols after adding the cyclic prefix CP have corresponding OFDM symbols affected by the channel (consisting of 256 complex numbers). The OFDM symbols affected by the channel are real signals, and the real signals are saved as a one-dimensional array with a shape of \((100000\times256)\), and each number in the array is a complex number.

[0130] While generating the simulation signal in the form of 256 subcarriers above, perform IDFT on each OFDM symbol (the simulation signal formed by 256 subcarriers, consisting of 256 complex numbers) in the generated simulation signal and add the cyclic prefix CP, so that the work before OFDM enters the channel is completed; the signal after the above processing will then pass through the channel, and the channel will have two parts of influence on the signal: ① Convolve the signal with the channel impulse response, ② Add noise to the convolved signal. During convolution, use the channel matrix \(H\) (with a shape of \(9000\times126\times2\), 9000 is the number of channel slices, 126 is the time delay, 2 is for two paths of IQ, and each of the 9000 pieces of data is a channel impulse response) collected and processed in a certain environment. For each of the 100,000 OFDM simulation symbols, each time randomly select a group of channel data from 9000 groups of channel slices, and then convolve this one OFDM symbol (the 256 complex numbers corresponding to 256 subcarriers) with this group of channel data; this is to fully cover and affect the 9000 groups of channel data. After completing the channel convolution, add complex Gaussian noise according to the set signal-to-noise ratio (SNR). Finally, at the receiving end, remove the cyclic prefix CP and perform the fast Fourier transform DFT, so that an OFDM frequency-domain symbol affected by the channel (consisting of 256 complex numbers) is obtained.

[0131] The above operations need to be performed for each of the 100,000 OFDM simulation symbols. Therefore, like the simulation signal, the real signals of each modulation method are finally saved as a one-dimensional array with a shape of \((100000\times256)\), and each number in the array is a complex number. The real signals of each saved modulation method are also frequency-domain signals, and finally, the real signals of 12 modulation methods are saved in total.

[0132] Other steps and parameters are the same as those in any one of the specific embodiments one to four.

[0133] Specific embodiment six: The difference between this embodiment and any one of the specific embodiments one to five is that in step two, the simulation signal and the real signal obtained in step one are subjected to normalization processing to obtain the normalized simulation signal and the normalized real signal;

[0134] The specific process is as follows:

[0135] Convert the simulated signal and the real signal obtained in Step 1 into a distribution with a mean of 0 and a standard deviation of 1.

[0136] Introduction to the signal standardization process

[0137] In diffusion, in order for the diffusion model to better learn the mapping relationship between the simulated signal and the real signal, after loading the simulated signals and real signals of various modulation signals saved, we perform the same standard standardization processing on these signals. Specifically, during the training process, we use the simulated signal for training as the standard, calculate the mean and standard deviation, and at the same time standardize and save the simulated signal data using this standard; at the same time, the corresponding real signal is also standardized using the same distribution benchmark of the above simulated signal. Finally, the standardized signal data is converted into a distribution with a mean of 0 and a standard deviation of 1. The purpose of standardization is to facilitate the model to better learn the feature relationship and improve the training efficiency and model performance.

[0138] Since the data flowing in the diffusion model is all processed after standardization and then trained, therefore, after the model finally generates a signal by reverse prediction, we need to perform inverse standardization on the generated signal and then save it, that is, use the data standardization benchmark saved before training to re-inverse standardize the generated signal data, convert it into the original style of the signal data, and finally restore and save it in the same format as the original signal data. The shape of the signal data before and after standardization does not change. It is only for improving the training efficiency and performance of the model that the distribution of the signal data becomes a standardized distribution.

[0139] Other steps and parameters are the same as those in any one of the first to fifth specific embodiments.

[0140] Specific Embodiment 7: The difference between this embodiment and any one of the first to sixth specific embodiments is that in Step 3, datasets for PSK group modulation mode, QAM group modulation mode, and APSK group modulation mode are obtained based on the real signal after standardization processing in Step 2;

[0141] The specific process is as follows:

[0142] Group the real signals after standardization processing in Step 2 according to the modulation mode, into PSK group real signals, QAM group real signals, and APSK group real signals;

[0143] For the PSK group modulation mode, the datasets are BPSK real signals, QPSK real signals, 8PSK real signals, and 16PSK real signals;

[0144] For the QAM group modulation method, the data sets are 16QAM real signals and 64QAM real signals;

[0145] For the APSK group modulation method, the data sets are 16APSK real signals, 32APSK real signals, and 64APSK real signals.

[0146] Other steps and parameters are the same as those in any one of the specific embodiments 1 to 6.

[0147] Specific embodiment 8: The difference between this embodiment and any one of the specific embodiments 1 to 7 is that in step 41

[0148] Perform forward diffusion on the signal data in the data set of the PSK group modulation method to obtain the signal data after forward diffusion, and save the signal data x t , t, and the actually added noise ε;

[0149] Perform forward diffusion on the signal data in the data set of the QAM group modulation method to obtain the signal data after forward diffusion, and save the signal data x t , t, and the actually added noise ε;

[0150] Perform forward diffusion on the signal data in the data set of the APSK group modulation method to obtain the signal data after forward diffusion, and save the signal data x t , t, and the actually added noise ε;

[0151] The specific process is as follows:

[0152] 1), Perform forward diffusion on the signal data in the data set of the PSK group modulation method to obtain the signal data after forward diffusion, and save the signal data x t , t, and the actually added noise ε;

[0153] The specific process of forward diffusion is as follows:

[0154]

[0155] Among them, α t = 1 - β t , ε ~ N(0, I);

[0156] x0 represents the BPSK real signal, QPSK real signal, 8PSK real signal, and 16PSK real signal data in the data set; x t represents the data after adding t steps of noise step by step from the original data x0;

[0157] α trepresents the attenuation rate of the true signal at time step t;

[0158] α i represents the attenuation rate of the true signal at time step i;

[0159] represents α i 's cumulative product;

[0160] ε represents noise, and N(0, I) represents a normal distribution with a mean of 0 and a covariance of the identity matrix;

[0161] β t represents the noise intensity that gradually increases with the increase of time steps;

[0162] I represents the identity matrix;

[0163] t represents the time step, and the range is t = 0, 1, …, T;

[0164] When reaching the last time step T, x T will become signal data full of noise;

[0165] 2), perform forward diffusion on the signal data in the dataset corresponding to the QAM modulation method, obtain the signal data after forward diffusion, and save the signal data x t , t, and the actually added noise ε;

[0166] The specific process of forward diffusion is as follows:

[0167]

[0168] where, α t = 1 - β t , ε ∼ N(0, I);

[0169] x0 represents the 16QAM true signal and 64QAM true signal data in the dataset; x t represents the data after adding t steps of noise step by step from the original data x0 according to the time step;

[0170] α t represents the attenuation rate of the true signal at time step t;

[0171] α i represents the attenuation rate of the true signal at time step i;

[0172] represents α i 's cumulative product;

[0173] ε represents noise, and N(0, I) represents a normal distribution with a mean of 0 and a covariance of the identity matrix;

[0174] β t represents the noise intensity that gradually increases with the increase of time steps;

[0175] I represents the identity matrix;

[0176] t represents the time step, and the range is t = 0, 1, …, T;

[0177] When reaching the last time step T, x T will become signal data that is all noise;

[0178] 3), perform forward diffusion on the signal data in the dataset corresponding to the APSK group modulation method to obtain the signal data after forward diffusion, and save the signal data x t , t, and the actually added noise ε;

[0179] The specific process of forward diffusion is as follows:

[0180]

[0181] Among them, α t = 1 - β t , ε ∼ N(0, I);

[0182] x0 represents the 16APSK true signal, 32APSK true signal, and 64APSK true signal data in the dataset; x t represents the data after adding t steps of noise step by step from the original data x0;

[0183] α t represents the attenuation rate of the true signal at time step t;

[0184] α i represents the attenuation rate of the true signal at time step i;

[0185] represents the i cumulative product of α;

[0186] ε represents the noise, and N(0, I) represents the normal distribution with a mean of 0 and a covariance of the identity matrix;

[0187] β t represents the noise intensity that gradually increases with the increase of time steps;

[0188] I represents the identity matrix;

[0189] t represents the time step, and the range is t = 0, 1, …, T;

[0190] When reaching the last time step T, xT The signal data will become all noise.

[0191] Forward diffusion process:

[0192] Some pairs of simulation signals and real signals of different modulation types are needed for training the diffusion model. The simulation signals mentioned here are the signals before passing through the channel environment, and the real signals are the signals after passing through a certain channel environment, such as Figure 2 , which shows the simulation signals and real signals of 32APSK.

[0193] After generating and saving the simulation signals and real signals for training the diffusion model, we use the diffusion model to learn the mapping relationship between the two sets of signals, so as to generate the real signals of the new modulation method;

[0194] The overall process of the Diffusion model is divided into two parts: the forward diffusion process and the reverse denoising process; we will first introduce the specific implementation of the forward diffusion process according to the experimental steps;

[0195] At present, the key applications of the diffusion model are in the processing and generation of images. Usually, the shape of a picture data is three-dimensional (H×W×C). Since the signal data we use are all I / Q signals, we analogize each signal sample to be processed to the picture format here.

[0196] According to the characteristics of the diffusion model, in order to ensure that it can be generalized to the generation of signals of new modulation methods, different diffusion models are trained separately according to different modulation types, such as: PSK group, QAM group, etc.; in each group, in order to fully train the diffusion model, we plan to use multiple signals to train the model and finally generate one signal. For example, for the PSK modulation method, the diffusion training set is BPSK, QPSK, 8PSK, 16PSK, then the finally generated signal is 32PSK;

[0197] The forward diffusion process of the signal data can be expressed by the following formula:

[0198]

[0199] where t represents the time step reached, and the range is 0 to T; x t represents the data after adding t steps of noise step by step from the original data x0; β t represents the noise intensity that gradually increases with the increase of the time step; I represents the added Gaussian noise. This process has a total of T steps and will generate a series of noise samples x1, x2,..., x T。When reaching the last time step T, x T will become signal data that is almost entirely noise.

[0200] To avoid adding noise to the data iteratively and instead directly sample from the original data x0 to the data x at the specified time step t , we re-parameterize and derive the above formula, and finally obtain the forward diffusion process formula as follows:

[0201]

[0202] where, α t = 1 - β t , ε ~ N(0, I);

[0203] In this study, the value of T is set to 1000, β start = 0.0001, β end = 0.02.

[0204] According to the above formula, in the forward diffusion process, we use the true signal for training diffusion as x0, and to avoid the model overly focusing on earlier time steps due to a large loss in the early stage of training, for each training sample, we randomly select a time step t within (0, T); then apply the Gaussian noise ε corresponding to the time step t to the true signal sample x0 to obtain the signal data x after fusing a certain amount of noise t . After processing each training sample according to the above steps, in addition to saving x t , t and the actually added noise ε should also be saved for training the U-Net network to predict the added noise in the next step.

[0205] Other steps and parameters are the same as those in any one of the specific embodiments one to seven.

[0206] Specific Embodiment Nine: The difference between this embodiment and any one of the specific embodiments one to eight is that in step four two, the signal data x after forward diffusion obtained by the PSK group modulation method t , the time step t, and the simulation signal after the normalization process in step two are used as the input of the U-Net network, and the predicted noise ε added at the t time step θ is used as the output of the U-Net network. Based on the true noise ε saved in the forward diffusion process and the predicted noise ε added at the t time step θ calculate the loss function L until the loss function L converges to obtain the trained U-Net network for the PSK group modulation method;

[0207] The signal data x after forward diffusion obtained by the QAM group modulation method t, the simulation signal after the time step t and the normalization process in step two are used as the input of the U-Net network, and the predicted noise ε added at the t time step θ is used as the output of the U-Net network. Based on the true noise ε saved during the forward diffusion process and the predicted noise ε added at the t time step θ calculate the loss function L until the loss function L converges, and obtain the trained U-Net network for the QAM group modulation method;

[0208] Use the signal data x after forward diffusion obtained by the APSK group modulation method t , the time step t, and the simulation signal after the normalization process in step two as the input of the U-Net network, and the predicted noise ε added at the t time step θ is used as the output of the U-Net network. Based on the true noise ε saved during the forward diffusion process and the predicted noise ε added at the t time step θ calculate the loss function L until the loss function L converges, and obtain the trained U-Net network for the APSK group modulation method;

[0209] The specific process is as follows: 1)

[0211] Use the signal data x after forward diffusion obtained by the PSK group modulation method t and the BPSK simulation signal, QPSK simulation signal, 8PSK simulation signal, 16PSK simulation signal after the normalization process in step two are concatenated in the channel dimension to obtain the concatenated data x' t ;

[0212] Use the concatenated data x' t and the time step t as the input of the U-Net network, and the predicted noise ε added at the t time step θ is used as the output of the U-Net network;

[0213] Based on the true noise ε saved during the forward diffusion process and the predicted noise ε added at the t time step θ calculate the loss function L, expressed as:

[0214] L = ||ε - ε θ (x' t , t)|| 2

[0215] where ε θ (x' t means that the concatenated data x' t and the time step t are used as the input of the U-Net network, and the U-Net network outputs the predicted noise ε added at the t time step θ ;

[0216] x' t represents the concatenated data, and t represents the time step; || || 2 represents the square of the norm;

[0217] The parameters of the U-Net network are optimized using backpropagation, expressed as:

[0218]

[0219] where θ represents the weights of the U-Net network and η represents the learning rate, represents the weight gradient of backpropagation;

[0220] Until the loss function L converges, a trained U-Net network for the PSK group modulation method is obtained; 2)、

[0222] The forward-diffused signal data x obtained by the QAM group modulation method t and the 16QAM simulation signal and 64QAM simulation signal after the normalization process in step two are concatenated in the channel dimension to obtain the concatenated data x'; t ;

[0223] The concatenated data x' t and the time step t are used as the input of the U-Net network, and the noise ε added at the predicted t time step θ is used as the output of the U-Net network;

[0224] Based on the true noise ε saved during the forward diffusion process and the noise ε added at the predicted t time step θ the loss function L is calculated, expressed as:

[0225] L = ||ε - ε θ (x' t , t)|| 2

[0226] where ε θ (x' t , t) represents that the concatenated data x' t and the time step t are used as the input of the U-Net network, and the U-Net network outputs the noise ε added at the predicted t time step θ ;

[0227] x' t represents the concatenated data, and t represents the time step; || || 2 represents the square of the norm;

[0228] The parameters of the U-Net network are optimized using backpropagation, expressed as:

[0229]

[0230] Among them, θ represents the weights of the U-Net network, η represents the learning rate, represents the weight gradient of backpropagation;

[0231] Until the loss function L converges, a trained U-Net network for the QAM group modulation method is obtained; 3)

[0233] The forward-diffused signal data x obtained by the APSK group modulation method t and the 16APSK simulation signal, 32APSK simulation signal, and 64APSK simulation signal after the normalization process in step two are concatenated in the channel dimension to obtain the concatenated data x'; t ;

[0234] The concatenated data x' t and the time step t are used as the input of the U-Net network, and the predicted noise ε added at the t time step θ is used as the output of the U-Net network;

[0235] Based on the true noise ε saved during the forward diffusion process and the predicted noise ε added at the t time step θ The loss function L is calculated and expressed as:

[0236] L = ||ε - ε θ (x' t , t)|| 2

[0237] Among them, ε θ (x' t , t) represents that the concatenated data x' t and the time step t are used as the input of the U-Net network, and the U-Net network outputs the predicted noise ε added at the t time step θ ;

[0238] x' t represents the concatenated data, and t represents the time step; |||| 2 represents the square of the norm;

[0239] The parameters of the U-Net network are optimized by backpropagation, expressed as:

[0240]

[0241] Among them, θ represents the weights of the U-Net network, η represents the learning rate, represents the weight gradient of backpropagation;

[0242] Until the loss function L converges, a trained U-Net network for APSK group modulation is obtained.

[0243] According to the above description of the forward diffusion and noise prediction processes for individual signal samples, we trained different groups of diffusion models in batches and for a certain number of training epochs for different modulation methods, and saved the relevant model weights and parameters.

[0244] After generating the corresponding signal data x for each signal sample t we then proceed with the noise prediction process. The main purpose of this step is to, through x t and the simulated signal, pass through the U-Net network to finally be able to inversely predict the noise ε added at time t, which is used for the next step of reverse denoising.

[0245] The U-Net network is a convolutional neural network for image segmentation. Its structure can be generally summarized as a symmetric U-shaped architecture, divided into an encoding path and a decoding path. The encoding path gradually reduces the spatial dimension of the feature map through multiple layers of convolution and pooling. The decoding path gradually restores the feature map size through deconvolution and upsampling operations. In addition, in each upsampling step, the decoding path concatenates the feature maps of the corresponding layer in the encoding path. The U-Net network structure used in this study is as Figure 4 shown.

[0246] Other steps and parameters are the same as those in any one of the first to eighth specific embodiments.

[0247] Specific Embodiment Ten: The difference between this embodiment and any one of the first to ninth specific embodiments is that for the trained U-Net network based on the PSK modulation method in step 43, the standardized 32PSK simulated signal is input into the trained U-Net network for the PSK modulation method, and reverse denoising starts from the pure noise x T and passes through x T x T-1 x T-2 ... x1 to generate the final predicted value of the 32PSK true signal; the predicted value of the finally generated 32PSK true signal data is inverse standardized (the inverse standardization in step 2) to obtain the predicted value of the finally generated 32PSK true signal data after inverse standardization;

[0248] For the trained U-Net network based on the QAM modulation method, the standardized 256QAM simulated signal is input into the trained U-Net network for the QAM modulation method, and reverse denoising starts from the pure noise x T and passes through x T x T-1 x T-2, …, x1, generate the predicted value of the final 256QAM real signal; perform inverse normalization on the predicted value of the finally generated 256QAM real signal data (the inverse normalization in step 2) to obtain the predicted value of the finally generated 256QAM real signal data after inverse normalization;

[0249] Input the normalized 128APSK simulation signal into the trained U-Net network based on the APSK modulation method. Starting from the pure noise x T Perform reverse denoising. After passing through x T , x T-1 , x T-2 , …, x1, generate the predicted value of the final 128APSK real signal; perform inverse normalization on the predicted value of the finally generated 128APSK real signal data (the inverse normalization in step 2) to obtain the predicted value of the finally generated 128APSK real signal data after inverse normalization;

[0250] The specific process is as follows:

[0251] 1). Input the normalized 32PSK simulation signal into the trained U-Net network based on the PSK modulation method. Starting from the pure noise x T Perform reverse denoising. After passing through x T , x T-1 , x T-2 , …, x1, generate the predicted value of the final 32PSK real signal; perform inverse normalization on the predicted value of the finally generated 32PSK real signal data (the inverse normalization in step 2) to obtain the predicted value of the finally generated 32PSK real signal data after inverse normalization;

[0252] The specific process is as follows:

[0253] 11). Let the time step t = T;

[0254] 12). Input the pure noise signal x t after forward diffusion at time step t obtained for the PSK modulation method, the 32PSK simulation signal after normalization in step two, and the time step t into the trained U-Net network based on the PSK modulation method. The trained U-Net network based on the PSK modulation method outputs the predicted noise ε θ (x t , t);

[0255] x T represents the signal data at time step T (signal data that is all noise);

[0256] 13) The noise ε added to the predicted time step t output by the trained U-Net network based on the PSK modulation method θ (x t , t), to obtain the denoised signal μ θ (x t , t) corresponding to the time step t; The formula is as follows:

[0257]

[0258] 14) Determine whether the time step t is equal to 1;

[0259] If so, obtain the predicted value of the 32PSK true signal data

[0260] If not, let t = t - 1, and repeat steps 12) to 14) until the predicted value of the 32PSK true signal data is obtained

[0261]

[0262] Among them, represents the predicted value of the 32PSK true signal data obtained at the time step t - 1, represents the predicted value of the finally generated 32PSK true signal data;

[0263] Perform inverse normalization on the predicted value of the finally generated 32PSK true signal data (the inverse normalization in step 2) to obtain the predicted value of the finally generated 32PSK true signal data after inverse normalization;

[0264] Save the predicted value of the finally generated 32PSK true signal data after inverse normalization in the same format as the true signal in step 1.

[0265] 2) For the trained U-Net network based on the QAM modulation method, the standardized 256QAM simulation signal is input into the trained U-Net network with the QAM modulation method, and start reverse denoising from the pure noise x T and go through x T 、x T-1 、x T-2 、…、x1 to generate the final predicted value of the 256QAM true signal; Perform inverse normalization on the predicted value of the finally generated 256QAM true signal data (the inverse normalization in step 2) to obtain the predicted value of the finally generated 256QAM true signal data after inverse normalization;

[0266] The specific process is as follows:

[0267] 21) Let the time step t = T; 22)、

[0269] Forward-diffuse the pure noise signal x obtained for the QAM modulation scheme by the time step t T the 256QAM simulation signal after the normalization process in Step 2 and the time step t are input into the trained U-Net network for the QAM modulation scheme. The trained U-Net network for the QAM modulation scheme outputs the predicted noise ε added at the time step t θ (x t , t);

[0270] x T represents the signal data at the time step T (signal data that is all noise);

[0271] 23)、Based on the predicted noise ε added at the time step t output by the trained U-Net network for the QAM modulation scheme θ (x t , t), obtain the denoised signal μ corresponding to the time step t θ (x t , t); The formula is as follows:

[0272]

[0273] 24)、Judge whether the time step t is equal to 1;

[0274] If so, obtain the predicted value of the 256QAM true signal data

[0275] If not, let t = t - 1, and repeat steps 22) to 24) until the predicted value of the 256QAM true signal data is obtained

[0276]

[0277] where, represents the predicted value of the 256QAM true signal data obtained at the time step t - 1, represents the predicted value of the finally generated 256QAM true signal data;

[0278] Perform inverse normalization on the predicted value of the finally generated 256QAM true signal data (the inverse normalization of Step 2) to obtain the predicted value of the finally generated 256QAM true signal data after inverse normalization;

[0279] Save the predicted value of the finally generated 256QAM true signal data after inverse normalization in the same format as the true signal in Step 1.

[0280] 3), The trained U-Net network based on APSK modulation method, the standardized 128-APSK simulation signal is input into the trained U-Net network of PSK modulation method, starting from pure noise x T to perform reverse denoising, passing through x T , x T-1 , x T-2 , …, x1, to generate the predicted value of the final 128-APSK true signal; perform inverse standardization on the predicted value of the final generated 128-APSK true signal data (the inverse standardization in step 2) to obtain the predicted value of the final generated 128-APSK true signal data after inverse standardization;

[0281] The specific process is as follows:

[0282] 31), Let the time step t = T;

[0283] 32), Input the pure noise signal x T forward-diffused at time step t obtained for APSK modulation method, the standardized 128-APSK simulation signal in step two, and the time step t into the trained U-Net network of APSK modulation method. The trained U-Net network of APSK modulation method outputs the predicted noise ε θ (x t , t);

[0284] x T represents the signal data at time step T (signal data that is all noise);

[0285] 33), Based on the predicted noise ε θ (x t , t) output by the trained U-Net network of APSK modulation method, obtain the denoised signal μ θ (x t , t); The formula is as follows:

[0286]

[0287] 34), Determine whether the time step t is equal to 1;

[0288] If so, obtain the predicted value of the 128-APSK true signal data

[0289] If not, let t = t - 1, and repeat steps 32) to 34) until the predicted value of the 128-APSK true signal data is obtained

[0290]

[0291] Among them, represents the predicted value of the 128APSK true signal data obtained at time step t-1, represents the predicted value of the finally generated 128APSK true signal data;

[0292] Perform inverse normalization on the predicted value of the finally generated 128APSK true signal data (inverse normalization in step 2) to obtain the predicted value of the finally generated 128APSK true signal data after inverse normalization;

[0293] Save the predicted value of the finally generated 128APSK true signal data after inverse normalization in the same format as the true signal in step 1.

[0294] The finally obtained is the signal data sample finally generated by the diffusion model for reverse denoising under the simulation signal with the modulation signal being clean; after predicting all the signal samples of clean, perform inverse normalization according to the same standard as in data preprocessing, and finally save it in the same format as the true signal in the diffusion training set. Thus, we have completed the process of the diffusion model generating signals.

[0295] After completing the forward diffusion process, the noise prediction process, and training and saving the model parameters, we start from pure noise data, gradually reverse predict the noise through the trained model, and continuously remove and update the data of the predicted noise to generate the final signal; the overall process of the reverse denoising process is as Figure 5 shown;

[0296] The overall formula of the reverse denoising process is as follows:

[0297] p θ (x t-1 |x t ) = N(x t-1 ; μ θ (x t ,t), ε θ (x t ,t))

[0298] Other steps and parameters are the same as those in any one of the specific embodiments one to nine.

[0299] Experimental results

[0300] After generating the signal according to the above steps, we assign corresponding labels to the generated signal and the real signal used for training the diffusion according to the modulation method, and finally form a complete modulated signal dataset. Combining with the current modulation recognition scheme, we select a CNN network to construct a modulation recognizer, and divide the synthesized modulated signal dataset into a training set, a validation set and a test set for training and testing. Finally, a confusion matrix is drawn as Figure 6 shown, where the modulated signals generated by diffusion are 32PSK, 256QAM, 128APSK, and the other modulation methods are the modulation methods of the real signals used for training diffusion. The final accuracy on the test set is 0.95.

[0301] From the performance of the confusion matrix and the test accuracy, it can be considered that the diffusion model can better learn the characteristics of the channel environment we introduced, and can generate the signals with specific modulation methods under this environment that we need, and finally the classification effect using the modulation recognizer is good.

[0302] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and deformations according to the present invention, but these corresponding changes and deformations should all fall within the protection scope of the appended claims of the present invention.

Claims

1. A method for generating a modulated signal based on a diffusion model, characterized in that: The specific process of the method is as follows: Step 1: Obtain paired simulation signals and real signals; Step 2: Standardize the simulation signals and real signals obtained in Step 1 to obtain the standardized simulation signals and real signals; Step 3: Based on the real signals after standardization in Step 2, obtain datasets for PSK modulation modes, QAM modulation modes, and APSK modulation modes; Step 4: Based on the simulation signals after standardization in Step 2 and the datasets for PSK modulation modes, QAM modulation modes, and APSK modulation modes obtained in Step 3, use a diffusion model to generate different modulation signals; the specific process is as follows: Step 4-1 Perform forward diffusion on the signal data in the dataset with the PSK group modulation method to obtain the signal data after forward diffusion, and save the signal data x t after forward diffusion, the time step t, and the actually added noise ε; Perform forward diffusion on the signal data in the QAM group modulation mode to obtain the signal data after forward diffusion, and save the signal data x t , the time step t, and the actually added noise ε; Perform forward diffusion on the signal data in the APSK group modulation mode to obtain the signal data after forward diffusion, and save the signal data x t , the time step t, and the actually added noise ε; Step 4-2 The signal data x after forward diffusion obtained by the PSK group modulation method t , the time step t, and the simulated signal after the normalization process in step 2 are used as the input of the U-Net network, and the predicted noise ε added at the t time step θ is used as the output of the U-Net network. Based on the true noise ε saved during the forward diffusion process and the predicted noise ε added at the t time step θ the loss function L is calculated until the loss function L converges, and a trained U-Net network for the PSK group modulation method is obtained; The signal data x after forward diffusion obtained by the QAM group modulation method t , the time step t, and the simulated signal after the normalization process in step 2 are used as the input of the U-Net network, and the predicted noise ε added at the t time step θ is used as the output of the U-Net network. Based on the true noise ε saved during the forward diffusion process and the predicted noise ε added at the t time step θ calculate the loss function L until the loss function L converges to obtain a trained U-Net network for the QAM group modulation method; The signal data x after forward diffusion obtained by the APSK group modulation method t , the time step t, and the simulation signal after the normalization process in step 2 are used as the input of the U-Net network, and the predicted noise ε added at the t time step θ is used as the output of the U-Net network. Based on the true noise ε saved during the forward diffusion process and the predicted noise ε added at the t time step θ calculate the loss function L until the loss function L converges to obtain a trained U-Net network for the APSK group modulation method; Step 4-3 The trained U-Net network based on the PSK modulation method, and the normalized 32PSK simulation signal is input into the trained U-Net network of the PSK modulation method, starting from pure noise x T to perform reverse denoising. After passing through x T , x T-1 , x T-2 , …, x1, the predicted value of the final 32PSK true signal is generated; the predicted value of the final generated 32PSK true signal data is denormalized to obtain the predicted value of the final generated 32PSK true signal data after denormalization; The trained U-Net network based on QAM modulation, the standardized 256QAM simulation signal is input into the trained U-Net network of QAM modulation, starting from pure noise x T to perform reverse denoising, passing through x T , x T-1 , x T-2 , …, x1, to generate the predicted value of the final 256QAM true signal; perform inverse standardization on the predicted value of the finally generated 256QAM true signal data to obtain the predicted value of the finally generated 256QAM true signal data after inverse standardization; The trained U-Net network based on APSK modulation, the standardized 128APSK simulation signal is input into the trained U-Net network with APSK modulation, starting from pure noise x T to perform reverse denoising. After passing through x T , x T-1 , x T-2 , …, x1, the predicted value of the final 128APSK real signal is generated; the predicted value of the final generated 128APSK real signal data is inverse-standardized to obtain the predicted value of the final generated 128APSK real signal data after inverse-standardization.

2. The modulation signal generation method based on a diffusion model according to claim 1, wherein: In Step 1, the paired simulation signals and real signals are obtained; the specific process is as follows: Step 1-1: Obtain simulation signals under different modulation modes; Step 1-2: Obtain real signals corresponding to the simulation signals under different modulation modes.

3. The method for generating a modulation signal based on a diffusion model according to claim 2, wherein: In Step 1-1, the simulation signals under different modulation modes are obtained; the specific process is as follows: Step 1-1-1 Generate a random bit stream for the QPSK modulation mode. The bit stream consists of bits, which are data of 1 or 0, and the occurrence probabilities of 0 and 1 in the bit stream are the same; Generate a random bit stream for the BPSK modulation mode. The bit stream consists of bits, which are data of 1 or 0, and the occurrence probabilities of 0 and 1 in the bit stream are the same; Generate a random bit stream for the 8PSK modulation mode. The bit stream consists of bits, which are data of 1 or 0, and the occurrence probabilities of 0 and 1 in the bit stream are the same; Generate a random bit stream for the 16PSK modulation mode. The bit stream consists of bits, which are data of 1 or 0, and the occurrence probabilities of 0 and 1 in the bit stream are the same; Generate a random bit stream for the 32PSK modulation mode. The bit stream consists of bits, which are data of 1 or 0, and the occurrence probabilities of 0 and 1 in the bit stream are the same; Generate a random bit stream for the 16QAM modulation mode. The bit stream consists of bits, which are data of 1 or 0, and the occurrence probabilities of 0 and 1 in the bit stream are the same; Generate a random bit stream for the 64QAM modulation mode. The bit stream consists of bits, which are data of 1 or 0, and the occurrence probabilities of 0 and 1 in the bit stream are the same; Generate a random bit stream for the 256QAM modulation mode. The bit stream consists of bits, which are data of 1 or 0, and the occurrence probabilities of 0 and 1 in the bit stream are the same; Generate a random bit stream for the 16APSK modulation mode. The bit stream consists of bits, which are data of 1 or 0, and the occurrence probabilities of 0 and 1 in the bit stream are the same; Generate a random bit stream for the 32APSK modulation mode. The bit stream consists of bits, which are data of 1 or 0, and the occurrence probabilities of 0 and 1 in the bit stream are the same; Generate a random bit stream for the 64APSK modulation mode. The bit stream consists of bits, which are data of 1 or 0, and the occurrence probabilities of 0 and 1 in the bit stream are the same; Generate a random bit stream for the 128APSK modulation mode. The bit stream consists of bits, which are data of 1 or 0, and the occurrence probabilities of 0 and 1 in the bit stream are the same; Step 1-1-2 For the QPSK modulation method, every 2 bits are mapped to a QPSK complex number, and every 256 QPSK complex numbers generate one OFDM symbol. A total of 100,000 OFDM symbols need to be generated; For the BPSK modulation method, every 1 bit is mapped to a BPSK complex number, and every 256 BPSK complex numbers generate one OFDM symbol. A total of 100,000 OFDM symbols need to be generated; For the 8PSK modulation method, every 3 bits are mapped to an 8PSK complex number, and every 256 8PSK complex numbers generate one OFDM symbol. A total of 100,000 OFDM symbols need to be generated; For the 16PSK modulation method, every 4 bits are mapped to a 16PSK complex number, and every 256 16PSK complex numbers generate one OFDM symbol. A total of 100,000 OFDM symbols need to be generated; For the 32PSK modulation method, every 5 bits are mapped to a 32PSK complex number, and every 256 32PSK complex numbers generate one OFDM symbol. A total of 100,000 OFDM symbols need to be generated; For the 16QAM modulation method, every 4 bits are mapped to a 16QAM complex number, and every 256 16QAM complex numbers generate one OFDM symbol. A total of 100,000 OFDM symbols need to be generated; For the 64QAM modulation method, every 6 bits are mapped to a 64QAM complex number, and every 256 64QAM complex numbers generate one OFDM symbol. A total of 100,000 OFDM symbols need to be generated; For the 256QAM modulation method, every 8 bits are mapped to a 256QAM complex number, and every 256 256QAM complex numbers generate one OFDM symbol. A total of 100,000 OFDM symbols need to be generated; For the 16APSK modulation method, every 4 bits are mapped to a 16APSK complex number, and every 256 16APSK complex numbers generate one OFDM symbol. A total of 100,000 OFDM symbols need to be generated; For the 32APSK modulation method, every 5 bits are mapped to a 32APSK complex number, and every 256 32APSK complex numbers generate one OFDM symbol. A total of 100,000 OFDM symbols need to be generated; For the 64APSK modulation method, every 6 bits are mapped to a 64APSK complex number, and every 256 64APSK complex numbers generate one OFDM symbol. A total of 100,000 OFDM symbols need to be generated; For the 128APSK modulation method, every 7 bits are mapped to a 128APSK complex number, and every 256 128APSK complex numbers generate one OFDM symbol. A total of 100,000 OFDM symbols need to be generated; Step One - One - Three For the QPSK modulation method, splice the 100,000 OFDM symbols generated in Step One - Two and save them as a one - dimensional array with a shape of (100,000 × 256). The one - dimensional array with a shape of (100,000 × 256) is the simulation signal of the QPSK modulation method; For the BPSK modulation method, concatenate the 100,000 OFDM symbols generated in step 112 and save them as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal of the BPSK modulation method; For the 8PSK modulation method, concatenate the 100,000 OFDM symbols generated in step 112 and save them as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal of the 8PSK modulation method; For the 16PSK modulation method, concatenate the 100,000 OFDM symbols generated in step 112 and save them as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal of the 16PSK modulation method; For the 32PSK modulation method, concatenate the 100,000 OFDM symbols generated in step 112 and save them as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal of the 32PSK modulation method; For the 16QAM modulation method, concatenate the 100,000 OFDM symbols generated in step 112 and save them as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal of the 16QAM modulation method; For the 64QAM modulation method, concatenate the 100,000 OFDM symbols generated in step 112 and save them as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal of the 64QAM modulation method; For the 256QAM modulation method, concatenate the 100,000 OFDM symbols generated in step 112 and save them as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal of the 256QAM modulation method; For the 16APSK modulation method, concatenate the 100,000 OFDM symbols generated in step 112 and save them as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal of the 16APSK modulation method; For the 32APSK modulation method, concatenate the 100,000 OFDM symbols generated in step 112 and save them as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal of the 32APSK modulation method; For the 64APSK modulation method, the 100,000 OFDM symbols generated in step 112 are concatenated and saved as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal of the 64APSK modulation method; For the 128APSK modulation method, the 100,000 OFDM symbols generated in step 112 are concatenated and saved as a one-dimensional array with a shape of (100,000×256). The one-dimensional array with a shape of (100,000×256) is the simulation signal of the 128APSK modulation method.

4. A modulation signal generation method based on a diffusion model according to claim 3, characterized in that: In step 12, obtain the real signal corresponding to the simulation signal under different modulation methods; the specific process is: Step 121: For the QPSK modulation method, obtain the real signal corresponding to the simulation signal under the QPSK modulation method; Step 122: For the BPSK modulation method, obtain the real signal corresponding to the simulation signal under the BPSK modulation method; Step 123: For the 8PSK modulation method, obtain the real signal corresponding to the simulation signal under the 8PSK modulation method; Step 124: For the 16PSK modulation method, obtain the real signal corresponding to the simulation signal under the 16PSK modulation method; Step 125: For the 32PSK modulation method, obtain the real signal corresponding to the simulation signal under the 32PSK modulation method; Step 126: For the 16QAM modulation method, obtain the real signal corresponding to the simulation signal under the 16QAM modulation method; Step 127: For the 64QAM modulation method, obtain the real signal corresponding to the simulation signal under the 64QAM modulation method; Step 128: For the 256QAM modulation method, obtain the real signal corresponding to the simulation signal under the 256QAM modulation method; Step 129: For the 16APSK modulation method, obtain the real signal corresponding to the simulation signal under the 16APSK modulation method; Step 120: For the 32APSK modulation method, obtain the real signal corresponding to the simulation signal under the 32APSK modulation method; Step 121: For the 64APSK modulation method, obtain the real signal corresponding to the simulation signal under the 64APSK modulation method; Step 122: For the 128APSK modulation method, obtain the real signal corresponding to the simulation signal under the 128APSK modulation method.

5. A modulation signal generation method based on a diffusion model according to claim 4, characterized in that: In step 121, for the QPSK modulation method, obtain the real signal corresponding to the simulation signal under the QPSK modulation method; The specific process is: 1), Perform inverse discrete Fourier transform IDFT on each OFDM symbol in the simulation signal generated by the QPSK modulation method, and add a cyclic prefix CP to each OFDM symbol after the inverse discrete Fourier transform IDFT to obtain each OFDM symbol after adding the cyclic prefix CP; 2), Each OFDM symbol after adding the cyclic prefix CP passes through the channel to obtain the OFDM symbol affected by the channel; the specific process is: 21), Let i = 1 22) Convolve the \(i\)-th OFDM symbol after adding the cyclic prefix CP with the channel impulse response to obtain the convolved signal. The specific process is as follows: The open-source Htrain.mat dataset contains 9000 groups of data. The shape of each group of data is \(64\times126\times2\), where 64 is the number of channels, 126 is the delay, and 2 represents two-way IQ. Select the data of the first channel in each of the 9000 groups of data as the channel data H. The shape of the channel data H is \(9000\times126\times2\), where 126 is the delay and 2 represents two-way IQ. Perform convolution calculation on the \(i\)-th OFDM symbol and a randomly selected group of data \(126\times2\) from the 9000 groups. 23) Add Gaussian noise to the convolved signal at a signal-to-noise ratio of 70 dB to obtain the signal after adding Gaussian noise. 24) Remove the cyclic prefix CP from the signal after adding Gaussian noise, and perform a fast Fourier transform DFT on the signal after removing the cyclic prefix CP to obtain the OFDM symbol affected by the channel. 25) Let \(i = i + 1\), and repeat steps 22) to 24) until the corresponding OFDM symbols affected by the channel are obtained for all 100000 OFDM symbols after adding the cyclic prefix CP. The OFDM symbols affected by the channel are the real signals, and the real signals are saved as a one-dimensional array with a shape of \((100000\times256)\).

6. The method for generating a modulation signal based on a diffusion model according to claim 5, wherein: In step two, perform normalization processing on the simulated signal and the real signal obtained in step one to obtain the normalized simulated signal and the normalized real signal. The specific process is as follows: Convert the simulated signal and the real signal obtained in step one into a distribution with a mean of 0 and a standard deviation of 1.

7. A modulation signal generation method based on a diffusion model according to claim 6, characterized in that: In step three, obtain datasets for PSK group modulation, QAM group modulation, and APSK group modulation based on the real signal after normalization processing in step two. The specific process is as follows: Group the real signal after normalization processing in step two according to the modulation method into the real signal of the PSK group, the real signal of the QAM group, and the real signal of the APSK group. For the PSK group modulation method, the datasets are the real signal of BPSK, the real signal of QPSK, the real signal of 8PSK, and the real signal of 16PSK. For the QAM group modulation method, the datasets are the real signal of 16QAM and the real signal of 64QAM. For the APSK group modulation method, the datasets are the real signal of 16APSK, the real signal of 32APSK, and the real signal of 64APSK.

8. A method for generating a modulated signal based on a diffusion model according to claim 7, characterized in that: In Step 41, forward diffusion is performed on the signal data in the PSK group modulation mode to obtain the signal data after forward diffusion, and the signal data x t , t, and the actually added noise ε are saved; Perform forward diffusion on the signal data in the QAM group modulation mode to obtain the signal data after forward diffusion, and save the signal data x t , t, and the actually added noise ε; Perform forward diffusion on the signal data in the APSK group modulation mode to obtain the signal data after forward diffusion, and save the signal data x t , t, and the actually added noise ε; The specific process is as follows: 1) Forward diffuse the signal data in the dataset with the PSK group modulation method to obtain the signal data after forward diffusion, and save the signal data x t , t, and the actually added noise ε; The specific process of forward diffusion is as follows: where α t = 1 - β t , ε ~ N(0, I); x0 represents the data of BPSK true signal, QPSK true signal, 8PSK true signal, and 16PSK true signal in the dataset; x t represents the data after adding t-step noise step by step in time steps starting from the original data x0; α t represents the attenuation rate of the true signal at time step t; α i represents the attenuation rate of the true signal at time step i; Denote α i cumulative product of; \(\varepsilon\) represents noise, and \(N(0, I)\) represents a normal distribution with a mean of 0 and a covariance of the identity matrix. β t represents the noise intensity that gradually increases with the increase of time steps; I represents the identity matrix. t represents the time step, and the range is \(t = 0, 1, \ldots, T\). When reaching the last time step T, x T will become signal data that is all noise; 2) Forward-diffuse the signal data in the dataset corresponding to the QAM group modulation method to obtain the signal data after forward diffusion, and save the signal data x t , t, and the actually added noise ε; The specific process of forward diffusion is as follows: where α t = 1 - β t , ε ~ N(0, I); x0 represents the data of 16QAM true signal and 64QAM true signal in the dataset; x t represents the data after adding t-step noise step by step in time steps starting from the original data x0; α t represents the attenuation rate of the true signal at time step t; α i represents the attenuation rate of the true signal at time step i; Denote α i cumulative product of; \(\varepsilon\) represents noise, and \(N(0, I)\) represents a normal distribution with a mean of 0 and a covariance of the identity matrix. β t represents the noise intensity that gradually increases with the increase of time steps; I represents the identity matrix. t represents the time step, and the range is \(t = 0, 1, \ldots, T\). When reaching the last time step T, x T will become signal data that is all noise; 3) Forward-diffuse the signal data in the dataset corresponding to the APSK group modulation method to obtain the signal data after forward diffusion, and save the signal data x t , t, and the actually added noise ε; The specific process of forward diffusion is as follows: where α t = 1 - β t , ε ~ N(0, I); x0 represents the data of 16APSK true signals, 32APSK true signals, and 64APSK true signals in the dataset; x t represents the data after adding t-step noise step by step in time steps starting from the original data x0; α t represents the attenuation rate of the true signal at time step t; α i represents the attenuation rate of the true signal at time step i; Denote α i cumulative product of; \(\varepsilon\) represents noise, and \(N(0, I)\) represents a normal distribution with a mean of 0 and a covariance of the identity matrix. β t represents the noise intensity that gradually increases with the increase of time steps; I represents the identity matrix. t represents the time step, where the range is t = 0, 1, …, T; When reaching the last time step T, x T will become signal data that is all noise.

9. A modulation signal generation method based on a diffusion model according to claim 8, characterized in that: In Step 42, the forward-diffused signal data x obtained by the PSK group modulation method t , the time step t, and the simulated signal after the normalization process in Step 2 are used as the input of the U-Net network, and the predicted noise ε added at the t time step θ is used as the output of the U-Net network. Based on the true noise ε saved during the forward diffusion process and the predicted noise ε added at the t time step θ the loss function L is calculated until the loss function L converges, and a trained U-Net network for the PSK group modulation method is obtained; The signal data x after forward diffusion obtained by the QAM group modulation method t , the time step t, and the simulation signal after the normalization process in step two are used as the input of the U-Net network, and the predicted noise ε added at the t time step θ is used as the output of the U-Net network. Based on the true noise ε saved during the forward diffusion process and the predicted noise ε added at the t time step θ the loss function L is calculated until the loss function L converges, and a trained U-Net network for the QAM group modulation method is obtained; The signal data x after forward diffusion obtained by the APSK group modulation method t , the time step t, and the simulated signal after the normalization process in step two are used as the input of the U-Net network, and the predicted noise ε added at the t time step θ is used as the output of the U-Net network. Based on the true noise ε saved during the forward diffusion process and the predicted noise ε added at the t time step θ the loss function L is calculated until the loss function L converges, and a trained U-Net network for the APSK group modulation method is obtained; The specific process is as follows: 1)、 The signal data x after forward diffusion obtained by the PSK group modulation method t and the BPSK simulation signal, QPSK simulation signal, 8PSK simulation signal, and 16PSK simulation signal after the normalization process in step 2 are concatenated in the channel dimension to obtain the concatenated data x t ′; The spliced data x t ′ and the time step t are used as the input of the U-Net network, and the predicted noise ε θ added at time step t is used as the output of the U-Net network; Based on the true noise ε saved during the forward diffusion process and the noise ε added at the predicted t time step θ Calculate the loss function L, expressed as: L = ||ε - ε θ (x t ′, t)|| 2 Among them, ε θ (x t ′, t) represents the concatenated data x t ′ and the time step t as the input to the U-Net network, and the U-Net network outputs the predicted noise ε added at the t time step θ ; x t ' represents the concatenated data, and t represents the time step; |||| 2 represents the square of the norm; Backpropagation is used to optimize the U-Net network parameters, expressed as: where θ represents the weights of the U-Net network, and η represents the learning rate, representing the weight gradient of backpropagation; Until the loss function L converges, a trained U-Net network for the PSK group modulation method is obtained; 2)、 The signal data x after forward diffusion obtained by the QAM group modulation method t and the 16QAM simulation signal and 64QAM simulation signal after the normalization process in step two are concatenated in the channel dimension to obtain the concatenated data x t ′; The spliced data x′ t and the time step t are used as the input to the U-Net network, and the predicted noise ε added at the t-th time step θ is used as the output of the U-Net network; Based on the true noise ε saved during the forward diffusion process and the noise ε added at the predicted t time step θ Calculate the loss function L, expressed as: L = ||ε - εθ(x′ t , t)|| 2 Among them, ε θ (x t ′, t) represents the concatenated data x t ′ and the time step t as the input to the U-Net network, and the U-Net network outputs the predicted noise ε added at the t time step θ ; x t ' represents the concatenated data, and t represents the time step; |||| 2 represents the square of the norm; Backpropagation is used to optimize the U-Net network parameters, expressed as: where θ represents the weights of the U-Net network and η represents the learning rate, represents the weight gradient of backpropagation; Until the loss function L converges, a trained U-Net network for the QAM group modulation method is obtained; 3)、 The signal data x after forward diffusion obtained by the APSK group modulation method t and the 16APSK simulation signal, 32APSK simulation signal, and 64APSK simulation signal after the normalization process in step two are concatenated in the channel dimension to obtain the concatenated data x t ′; The spliced data x t ′ and the time step t are used as the input of the U-Net network, and the predicted noise ε θ added at the t-th time step is used as the output of the U-Net network; Based on the true noise ε saved during the forward diffusion process and the noise ε added at the predicted t time step θ Calculate the loss function L, expressed as: L = ||ε - ε θ (x t ′, t)|| 2 where, ε θ (x t ′, t) represents the concatenated data x t ′ and the time step t as the input to the U-Net network, and the U-Net network outputs the predicted noise ε added at the t-th time step θ ; x′ t represents the concatenated data, and t represents the time step; |||| 2 represents the square of the norm; Backpropagation is used to optimize the U-Net network parameters, expressed as: Among them, θ represents the weights of the U-Net network, and η represents the learning rate, representing the weight gradient of backpropagation; Until the loss function L converges, a trained U-Net network for the APSK group modulation method is obtained.

10. A method for generating a modulated signal based on a diffusion model according to claim 9, characterized in that: The trained U-Net network based on the PSK modulation method in step 43. The normalized 32PSK simulation signal is input into the trained U-Net network of the PSK modulation method, starting from pure noise x T to perform reverse denoising. After x T , x T-1 , x T-2 , …, x1, the predicted value of the final 32PSK true signal is generated; the predicted value of the finally generated 32PSK true signal data is inverse-normalized to obtain the predicted value of the finally generated 32PSK true signal data after inverse-normalization; The trained U-Net network based on QAM modulation, the 256QAM simulation signal after normalization is input into the trained U-Net network with QAM modulation, starting from pure noise x T for reverse denoising, through x T , x T-1 , x T-2 , …, x1, to generate the predicted value of the final 256QAM true signal; perform denormalization on the predicted value of the finally generated 256QAM true signal data to obtain the predicted value of the finally generated 256QAM true signal data after denormalization; The trained U-Net network based on APSK modulation, the standardized 128APSK simulation signal is input into the trained U-Net network of APSK modulation, starting from pure noise x T for reverse denoising, passing through x T , x T-1 , x T-2 , …, x1, to generate the predicted value of the final 128APSK true signal; perform inverse standardization on the predicted value of the finally generated 128APSK true signal data to obtain the predicted value of the finally generated 128APSK true signal data after inverse standardization; The specific process is as follows: 1), The trained U-Net network based on the PSK modulation method, and the normalized 32PSK simulation signal is input into the trained U-Net network of the PSK modulation method, starting from pure noise x T for reverse denoising, passing through x T , x T-1 , x T-2 , …, x1, to generate the predicted value of the final 32PSK true signal; perform denormalization on the predicted value of the finally generated 32PSK true signal data to obtain the predicted value of the finally generated 32PSK true signal data after denormalization; the specific process is as follows: 11) Let the time step t = T; 12), the pure noise signal x after forward diffusion of the time step t obtained for the PSK modulation method t , the 32PSK simulation signal after the normalization process in step two and the time step t are input into the trained U-Net network for the PSK modulation method, and the trained U-Net network for the PSK modulation method outputs the predicted noise ε added at the time step t θ (x t , t); 13) The noise ε added to the predicted time step t output by the trained U-Net network based on the PSK modulation method θ (x t , t), to obtain the denoised signal μ θ (x t , t); The formula is as follows: 14) Determine whether the time step t is equal to 1; If so, obtain the predicted value of the 32PSK true signal data Otherwise, let t = t - 1, and repeat steps 12) to 14) until the predicted value of the 32PSK true signal data is obtained. Among them, represents the predicted value of the 32PSK true signal data obtained at time step t - 1, represents the predicted value of the finally generated 32PSK true signal data; Denormalize the predicted value of the finally generated 32PSK true signal data to obtain the denormalized predicted value of the finally generated 32PSK true signal data; 2) The trained U-Net network based on the QAM modulation method, and the 256QAM simulation signal after normalization processing is input into the trained U-Net network with the QAM modulation method, starting from pure noise x T to perform reverse denoising, passing through x T , x T-1 , x T-2 , …, x1, to generate the predicted value of the final 256QAM true signal; perform denormalization on the predicted value of the finally generated 256QAM true signal data to obtain the predicted value of the finally generated 256QAM true signal data after denormalization; the specific process is as follows: 21) Let the time step t = T; 22)、 The pure noise signal x after forward diffusion at time step t obtained for the QAM modulation scheme T , the 256QAM simulation signal after the normalization process in step two, and the trained U-Net network for the QAM modulation scheme with the input at time step t. The trained U-Net network for the QAM modulation scheme outputs the predicted noise ε added at time step t θ (x t , t); 23) The noise ε added to the predicted time step t output by the trained U-Net network based on the QAM modulation method θ (x t , t), to obtain the denoised signal μ corresponding to the time step t θ (x t , t); The formula is as follows: 24) Determine whether the time step t is equal to 1; If so, obtain the predicted value of the 256QAM true signal data Otherwise, let t = t - 1, and repeat steps 22) to 24) until the predicted value of the 256QAM true signal data is obtained. Among them, represents the predicted value of the 256QAM true signal data obtained at time step t-1, represents the predicted value of the finally generated 256QAM true signal data; Denormalize the predicted value of the finally generated 256QAM true signal data to obtain the denormalized predicted value of the finally generated 256QAM true signal data; 3), The trained U-Net network based on APSK modulation method, the standardized 128APSK simulation signal is input into the trained U-Net network of PSK modulation method, starting from pure noise x T for reverse denoising, through x T , x T-1 , x T-2 , …, x1, to generate the predicted value of the final 128APSK true signal; perform inverse standardization on the predicted value of the finally generated 128APSK true signal data to obtain the predicted value of the finally generated 128APSK true signal data after inverse standardization; the specific process is as follows: 31) Let the time step t = T; 32) Forward-diffuse the pure noise signal x after the time step t obtained for the APSK modulation method T The 128APSK simulation signal after the normalization process in Step 2 and the time step t are input into the trained U-Net network for the APSK modulation method. The trained U-Net network for the APSK modulation method outputs the predicted noise ε added at the time step t θ (x t , t); 33) The noise ε added to the predicted time step t output by the trained U-Net network based on the APSK modulation method θ (x t , t), to obtain the denoised signal μ θ (x t , t); The formula is as follows: 34) Determine whether the time step t is equal to 1; If so, obtain the predicted value of the 128APSK true signal data If not, let t = t - 1, and repeat the execution of 32) to 34) until the predicted value of the 128APSK true signal data is obtained Among them, represents the predicted value of the 128APSK true signal data obtained at time step t - 1, represents the predicted value of the finally generated 128APSK true signal data; Denormalize the predicted value of the finally generated 128APSK true signal data to obtain the denormalized predicted value of the finally generated 128APSK true signal data.

Images