Modulation pattern recognition method based on generative neural network architecture search

Abstract

The invention discloses a modulation pattern recognition method based on generative neural network architecture search, and relates to the technical field of artificial intelligence and the field of wireless communication, and the method comprises the steps: training a modulation pattern recognition network architecture search framework to obtain a generative model, a generation guide predictor and a performance predictor; performing cross-data-set modulation pattern recognition network architecture search, and generating an optimal modulation pattern recognition network architecture by the generation model according to a given modulation pattern data set; performing parameter training on the optimal modulation pattern recognition network architecture according to a given modulation recognition data set to obtain a modulation pattern recognition network model; deploying the modulation pattern recognition model to channel transmission, and performing modulation pattern recognition on an electromagnetic signal transmitted based on the channel; according to the modulation pattern recognition method, the optimal modulation pattern recognition network model can be quickly and automatically searched according to the given modulation pattern data set, so that different application scenes can be quickly adapted. Meanwhile, according to the method, the search overhead can be greatly reduced in the network architecture search process, and the requirement for the data set scale is lowered.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a modulation pattern recognition method based on generative neural network architecture search. Background Art

[0002] In modern communication systems, signal modulation pattern recognition technology is an important part of communication quality assurance and spectrum resource management. Different types of signal modulation patterns are widely used in wireless transmission to meet different communication needs. These modulation patterns include but are not limited to common amplitude modulation (AM), frequency modulation (FM), phase modulation (PM) and their combinations, such as QAM (quadrature amplitude modulation), PSK (phase modulation), QPSK (quadrature phase modulation), etc.

[0003] With the increase in signal complexity, automatic identification of modulation types has become a key technical issue that needs to be solved urgently. In wireless communications, especially in the context of spectrum sharing and interference management, timely and accurate identification of signal modulation patterns is crucial to improving the system's anti-interference capability and optimizing spectrum resource utilization. Traditional modulation recognition methods rely on preset feature extraction and classification algorithms. Although they have achieved certain application results, the recognition accuracy and robustness of traditional methods are facing challenges as signal diversity and noise environments change.

[0004] With the continuous development of machine learning and deep learning technologies, data-driven modulation recognition methods have gradually become a hot topic of research. However, at present, such methods usually require manual design and adjustment of the neural network architecture, and the experience of designing the model architecture often comes from other fields such as images and vision, which leads to the network model not necessarily being able to adapt well to the modulation recognition problem of communication signals. In addition, the current recognition method still has poor recognition performance under low signal-to-noise ratio conditions. Therefore, how to automatically generate a reliable model architecture to adapt to the communication modulation recognition problem is a very promising research direction.

[0005] Neural Architecture Search aims to automate the network architecture design process and discover effective structures of neural networks for specific problems while minimizing human intervention. Early neural network architecture search techniques were mainly implemented through reinforcement learning and evolutionary algorithms. The search strategy was updated by evaluating the searched network architecture, which consumed huge computing resources and could not be searched on large search spaces and data sets. Differentiable architecture search speeds up the search, but the network architecture selected from the search space usually falls into a suboptimal state. Therefore, there is a technology that models the search space through generative models and directly generates architectures from promising areas. However, the above methods have a common problem, that is, the search algorithm needs to be retrained on each data set, which will cause a lot of resource waste when searching on multiple data sets. Summary of the invention

[0006] Based on the technical problems existing in the background technology, the present invention proposes a modulation style recognition method based on generative neural network architecture search, which greatly reduces the search overhead and reduces the dependence on the amount of data in the target data set.

[0007] The present invention proposes a modulation pattern recognition method based on generative neural network architecture search; comprising: Train a modulation style recognition network architecture search framework to obtain a generative model, a generative guidance predictor, and a performance predictor; Guide the search for modulation style recognition network architecture across data sets, and generate a model to generate the optimal modulation style recognition network architecture based on a given modulation style data set; According to the given modulation recognition data set, the parameters of the optimal modulation pattern recognition network architecture are trained to obtain the modulation pattern recognition network model; The modulation pattern recognition model is deployed on the channel transmission to perform modulation pattern recognition on the electromagnetic signal transmitted through the channel.

[0008] Furthermore, the training process of the modulation style recognition network architecture search framework is as follows: A parameterized search space of a modulation style recognition network architecture is constructed based on the modulation style recognition task, wherein each modulation style recognition network architecture in the parameterized search space is regarded as a directed acyclic graph, wherein the nodes of the graph represent operations, and the directed edges represent the direction of information transmission, and each style recognition network architecture is represented as an adjacency matrix and an operator operation type matrix; The network architecture encoder-decoder is trained in the parameterized search space, and the modulation style recognition network architecture is encoded into the latent space through the network architecture encoder-decoder; A diffusion model is trained in latent space, and the reverse diffusion process of the diffusion model is used as a generative model of the modulation style recognition network architecture, a performance predictor of the modulation style recognition network architecture, and a generation guidance predictor of the modulation style recognition network architecture.

[0009] Furthermore, the network architecture codec is trained in the parameterized search space of the modulation style recognition network architecture as follows: Noise is added to the modulation style recognition network architecture representation matrix to train the architecture encoder to encode the modulation style recognition network architecture into a potential representation vector, which is input into the architecture decoder after adding noise, and a reconstruction loss is constructed based on the output modulation style recognition network architecture and the input modulation style recognition network architecture to train the architecture encoder-decoder.

[0010] Furthermore, the diffusion model is trained in the latent space, specifically: A stochastic differential equation is defined to describe the unconditional forward diffusion process of the modulation style recognition network architecture. The modulation style recognition network architecture in the parameterized search space is gradually noised through the forward diffusion process, and the distribution of the modulation style recognition network architecture is mapped to the known prior distribution to obtain the hidden representation of the architecture after noise addition. Define an inverse stochastic differential equation to describe the unconditional diffusion process of the modulation style recognition network architecture, and denoise the hidden representation of the noisy architecture through the reverse diffusion process; In back-diffusion, the trainable parameters of the scoring network are trained to approximate the unknown term score function in the back-diffusion process, thereby obtaining a trained generative model.

[0011] Furthermore, 5-1) a generative guidance predictor of a modulation style recognition network architecture is trained in latent space, specifically: Build a cross-scenario dataset, including sub-datasets, modulation style recognition network architecture, and true accuracy triples; Input the modulation style recognition network architecture representation matrix into the architecture encoder for encoding to obtain the latent space representation vector of the architecture; According to the forward diffusion process, choose the time , diffuse and add noise to the latent space vector; The sub-dataset and the latent space vector after noise are input into the guided predictor to predict the corresponding true accuracy. The mean square error loss training is constructed based on the predicted accuracy and the true accuracy to generate the guided predictor model. During the deployment and execution of the modulation style recognition model, the architecture inputs of the generated guided predictor are the intermediate results of the diffusion model denoising, the denoised time steps, and the target dataset features; 5-2) Train the performance predictor of the modulation style recognition network architecture, specifically: Build a cross-scenario dataset, including sub-datasets, modulation style recognition network architecture, and true accuracy triples; Input the sub-dataset and the modulation style recognition network architecture representation matrix into the performance predictor, predict the corresponding true accuracy, and build a mean square error loss training performance predictor model based on the predicted accuracy and true accuracy; During the deployment and execution of the modulation style recognition model, the input of the performance predictor is the modulation style recognition network architecture finally generated by the diffusion model denoising and the characteristics of the target dataset.

[0012] Furthermore, the optimization expression of the trainable parameters in the generative model is as follows:

[0013] in, are the trainable parameters after the scoring network is trained, are the trainable parameters of the scoring network, represents the time of the diffusion process, Indicates expectation, Indicates time-based Given a positive weighting function, is the 0th moment, i.e. the function mapping of the initial modulation style recognition network architecture, For the moment Function mapping of the modulation pattern recognition network architecture, Indicates time 0 expectations, Indicates that in a given In case of moment hour expectations, Indicates and time is the input score network, is the unknown term score function, For about The gradient of Based on time of The marginal probability distribution of .

[0014] Furthermore, the cross-scenario dataset is constructed as follows: Preprocess the data set used for wireless communication modulation recognition to obtain the source data set, randomly sample k categories from the source data set each time to form a sub-data set, randomly sample a modulation style recognition network architecture in the parameterized search space, train the modulation style recognition network architecture to obtain the true accuracy on the sub-data set, and use the task triple consisting of the sub-data set, modulation style recognition network architecture, and true accuracy as cross-training data to construct a cross-scenario data set; A mean square error loss is constructed based on the sub-datasets and modulation style recognition network architecture in the cross-scene dataset to train a cross-domain predictor to search the modulation style recognition network architecture of the source dataset. The generative model generates the optimal modulation style recognition network architecture according to the task dataset. The cross-domain predictor includes a generation guidance predictor and a performance predictor.

[0015] Furthermore, the construction process of the source dataset is as follows: Dynamic simulation channel: Introduce Doppler effect and multipath effect to simulate the impact of real channel on signal and construct dynamic simulation channel: Data enhancement: The IQ signals in the data set are synthesized into an input complex signal. After the input complex signal passes through the dynamic simulation channel, the enhanced signal data is obtained. The K factor and Doppler frequency shift parameters of the channel are updated every several sampling points to realize the dynamic change of the channel. Filtering: The enhanced signal data is filtered based on a low-pass filter, and the processed data is used as the source data set; The expression of dynamic simulation channel is as follows:

[0016] in, for Signal data after time enhancement, is the Rice factor, is the Doppler frequency shift parameter, for The input signal at time, for The input signal at time, is the path gain, is the number of scattering paths, It is The scattering path gain, It is The delay under the scattering path is Is an imaginary unit.

[0017] Furthermore, both the generation guide predictor and the performance predictor are provided with a data set encoder, and the structures of the two data set encoders are consistent; For generating a guided predictor, the mean square error loss constructed by the predicted accuracy and the true accuracy is as follows:

[0018] in, Indicates time, The modulation style recognition network architecture is represented by the modulation style recognition network architecture after the architecture encoder, and then after time The result of forward diffusion and noise addition is is the sub-dataset obtained by sampling the source data set, For prediction accuracy, Identifying network architectures for modulation patterns exist The true accuracy of For parameters Regarding the task triplet The loss function is The task triple , For the distribution of tasks, represents the trainable parameters for generating the guided predictor; For the performance predictor, the mean square error loss constructed by the predicted accuracy and the true accuracy is as follows:

[0019] in, is the prediction accuracy, For parameters Regarding the task triplet The loss function is Representation performance predictor The trainable parameters of .

[0020] Furthermore, a guidance predictor and a performance predictor are generated by training a cross-scenario dataset. The generated guidance predictor is used to guide the score network to generate a latent space representation of the modulation style recognition network architecture through a diffusion model in the latent space. The architecture decoder maps the latent space representation generated by the diffusion model to the modulation style recognition network architecture in the original search space. The performance predictor is used to predict the performance of the mapped modulation style recognition network architecture. During the deployment and execution of the modulation style recognition network architecture search framework, a generation guided predictor guides the generation model to generate a set of modulation style recognition network architectures applied to a given modulation recognition data set, the modulation style recognition network architecture set is sorted in descending order based on the performance predictor, and the modulation style recognition network architecture with the highest performance ranking is selected for performance evaluation. The selected modulation style recognition network architecture and the corresponding evaluation accuracy are used to fine-tune the generation guided predictor and the performance predictor to adapt to the generation of a given modulation recognition data set.

[0021] The modulation style recognition method based on generative neural network architecture search provided by the present invention has the advantages that: based on the neural network architecture search technology, the network architecture most suitable for the modulation style recognition task is automatically searched from the data set, thereby avoiding the manual intervention in the traditional method, thereby improving the efficiency and accuracy of the recognition process. At the same time, in order to overcome the need for the existing network architecture search algorithm to be re-run on each data set, the idea of ​​the predictor-guided diffusion model generation method is used to encode the neural network architecture into the latent space, and the diffusion model is guided by the pre-trained predictor to generate the neural network architecture of the cross-scenario data set; when the modulation style recognition network architecture search framework is trained and deployed to an unknown data set, the modulation style recognition network architecture search framework does not need to be retrained, but can directly use the target data set features to generate the network architecture, greatly reducing the search overhead and reducing the dependence on the number of data in the target data set features. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 This is the spectrum of the Butterworth low-pass filter; Figure 3 It is the constellation diagram, amplitude and phase diagram, and spectrum diagram of the original signal and the denoised signal; Figure 4 This is a flowchart of the process of generating an excellent neural network architecture; Figure 5 is a schematic diagram of the architecture encoder-decoder in the latent space; Figure 6 Schematic diagram of the forward and reverse diffusion process of the diffusion model; Figure 7 It is a diagram of the training and deployment of predictors for cross-scenario datasets; Figure 8 This is a schematic diagram of a neural network architecture in NAS-Bench-201 and the corresponding adjacency matrix representation and node operation matrix representation diagram; Fig. 9 This is a schematic diagram of the search space in NAS-Bench-201.

[0023] Fig.10 It is a schematic diagram of the multi-head attention module structure; Fig.11 This is a schematic diagram of the attention block structure. DETAILED DESCRIPTION

[0024] Below, the technical solution of the present invention is described in detail through specific embodiments. Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific implementation disclosed below.

[0025] like Figures 1 to 11 As shown, the present invention proposes a modulation pattern recognition method based on generative neural network architecture search, which realizes rapid deployment on a target data set by training a predictor across data sets, and realizes rapid search for an excellent neural network architecture for an unknown task, wherein the target data set is a target data set composed of electromagnetic signals transmitted through a channel, and specifically includes the following steps 1 to 4: Step 1: training a modulation pattern recognition network architecture search framework to obtain a modulation pattern recognition network architecture generation model, a modulation pattern recognition network architecture performance predictor, and a modulation pattern recognition network architecture generation guidance predictor; The training process of the modulation style recognition network architecture search framework is (a1) to (a7): The details are as follows: (a1) Constructing a parameterized search space for the modulation style recognition network architecture based on the modulation style recognition task; The parameterized search space is a set of all possible neural network architectures. Each neural network architecture in the parameterized search space can be viewed as a directed acyclic graph, where the nodes of the graph represent operations and the directed edges represent the direction of information transfer. A network structure with N nodes and F predefined operations , by its operator type matrix and the upper triangular adjacency matrix Defined as . An adjacency matrix can be used to represent the connection mode of the neural network architecture, and an operation matrix can be used to represent the operation of the nodes in the neural network architecture. In this way, each neural network architecture in the parameterized search space can be represented by two matrices. Figure 8 This is a schematic diagram of a neural network architecture in NAS-Bench-201 and the corresponding adjacency matrix representation and node operation matrix representation. In the original search space Fig. 9 In , edges are used to represent operations, and nodes are used to represent the direction of calculation flow. The two can be transformed into each other, that is, each style recognition network architecture can be represented by two matrices: the adjacency matrix and the operator operation type matrix.

[0026] (a2) Train the network architecture encoder-decoder in the parameterized search space and encode the modulation style recognition network architecture into the latent space; After the representation of the modulation style recognition network architecture is determined, the generation problem of the modulation style recognition network architecture is transformed into the generation problem of two matrices. Only the corresponding matrices need to be generated to obtain the corresponding modulation style recognition network architecture. Considering that the adjacency matrix and the operator operation type matrix are both 0-1 discrete matrices, the Hamming distance between networks is at least 1, which is not conducive to the diffusion model learning an effective network coding distribution. Figure 5 As shown in the figure, according to the latent diffusion model technology, the architecture encoder and decoder are used to encode the modulation style recognition network architecture representation matrix into the latent space, and the latent space representation is used to achieve efficient and high-quality modulation style recognition network architecture generation. First, the architecture encoder and decoder encodes the modulation style recognition network architecture into a latent representation, and then the diffusion process is performed on this compact representation space. This strategy can further reduce the demand for computing resources, allowing the diffusion model to better model the search space, while improving the quality of architecture generation.

[0027] The input of the architecture encoder is the representation matrix of the modulation style recognition network architecture, that is, the adjacency matrix and the operator operation type matrix. The output result is the corresponding encoded latent vector of the latent space. The input of the architecture decoder is the encoded latent vector of the latent space, and the output is the reconstructed prediction of the representation matrix of the modulation style recognition network architecture corresponding to the latent vector. In the training phase, the encoder and decoder are connected together for joint training.

[0028] Taking the noise-based encoder-decoder network as an example, its architecture is mainly composed of stacked blocks consisting of convolution, LeakyReLU activation and InstanceNorm1d normalization, and the architecture encoder and decoder each have 4 blocks. The representation matrix of the modulation style recognition network architecture is converted into a one-dimensional vector and then input into the matrix for encoding. Gaussian noise with a standard deviation of 0.1 is added to the input to train the robustness of the architecture encoder to the input. Gaussian noise with a standard deviation of 0.1 is added to the intermediate encoding latent space to achieve distance distinction between different architectures, and the architecture encoder-decoder is trained using reconstruction loss.

[0029] (a3) training a diffusion model in latent space, and using the reverse diffusion process of the diffusion model as a generative model of the modulation style recognition network architecture, a performance predictor of the modulation style recognition network architecture, and a generation guidance predictor of the modulation style recognition network architecture; First, a stochastic differential equation (SDE) is defined to describe the unconditional diffusion process of the modulation style recognition network architecture. This process describes the process of starting from the modulation style recognition network architecture distribution in the search space, gradually adding noise perturbations and transforming to the prior noise distribution. Then, the perturbation process can be reversed and the structure of the search space can be sampled starting from the noise. The forward process maps the modulation style recognition network architecture distribution P(A0) to the known prior distribution P(A T ) is defined as the following ItôSDE: (1); in, is the drift factor, is the diffusion factor, Represents a cross-domain predictor The trainable parameters of are the standard Wiener process, is an infinitesimal white noise, For the moment The solution to this SDE is a continuous set of random variables that tracks the random trajectory of the process from time t0 to T. Subsequently, in the forward diffusion process, the architecture is perturbed with Gaussian noise at each step.

[0030] Every SDE has a corresponding inverse SDE. To sample from the distribution, the reverse time diffusion process corresponding to the forward process is modeled by the following SDE: (2); in, Indicates time-based of The marginal probability distribution of represents an infinitesimal negative time step, is the reverse-time standard Wiener process, For gradient.

[0031] In order to use the inverse process as a generative model, it is necessary to train a scoring network Trainable parameters of (score-net) , to fractional function of unknown terms in the reverse process Approximately, is a positive weighted function given by: The optimization expression of the trainable parameters in the scoring network is as follows: (3); in, are the trainable parameters after the scoring network is trained, Indicates expectation, is the 0th moment, i.e. the function mapping of the initial modulation style recognition network architecture, For the moment Function mapping of the modulation pattern recognition network architecture, Indicates time 0 expectations, Indicates that in a given In case of moment hour expectations, Indicates and time is the input score network, For about The gradient of Based on time of The marginal probability distribution of .

[0032] The input of the scoring network is the latent vector of the representation matrix of the modulation style recognition network architecture after being encoded by the encoder, and then after time The perturbation result obtained by the forward diffusion perturbation, and the time of the latent vector perturbation , the output is The predicted fit of , since Gaussian noise is added in the forward diffusion, the expected mean of the forward diffusion result is the same as the input, The form of can be mathematically derived according to the choice of diffusion process model. Taking the variance explosion type stochastic differential equation as an example, the upper triangular part of the random standard Gaussian noise is set to -1, and then this matrix plus its own transpose is obtained to obtain the matrix , The standard deviation of the moment diffusion process is std, and the perturbation result of the scoring network input is is a matrix Multiply by the standard deviation std, and then add the encoded latent vector. The loss function value is the score network output Multiply The standard deviation std of the diffusion process at each moment, and then add the result , and then take the square to get the final loss value.

[0033] It can be understood that the diffusion model is a mathematical model, and the score network is the network used to fit the unknown term in the diffusion model formula. The training set of the score network is all the modulation style recognition network architectures in the search space after being encoded by the architecture encoder. The input is the random time step. (i.e. time ) after adding noise to the neural network architecture representation , the target output is , this embodiment hopes that the score network can remove noise from the noisy representation. Once the score network is trained, the inverse process can be used to generate The modulation pattern recognition network architecture of . The noise can be sampled from a known prior distribution (Gaussian distribution) using Fit the unknown terms in the reverse process stochastic differential equation, simulate the opposite process to remove the noise step by step to obtain the network representation vector.

[0034] For the architecture of the diffusion model, the diffusion process can be used based on variance explosion SDE. Taking the transformation modulation style recognition network architecture with cross-attention calculation perception as an example, the scoring network has 12 layers, and each layer has 8 attention heads. The dimensions of all feedforward layers are , the size of the hidden layer is . Set the SDE , , discrete time step .

[0035] (a4) Preprocess the data set used for wireless communication modulation recognition to obtain the source data set. Construct a cross-scenario data set based on the preprocessed data set, randomly sample k categories from the source data set each time to form a sub-data set, randomly sample a modulation style recognition network architecture in the parameterized search space, train to obtain the true accuracy of the modulation style recognition network architecture on the sub-data set, and use the task triple consisting of the sub-data set, modulation style recognition network architecture, and true accuracy as cross-training data to construct a cross-scenario data set; The DeepSig RadioML 2018.01A dataset is used as an example to introduce the preprocessing method of the dataset. This dataset is a dataset widely used in automatic modulation recognition research and contains data of multiple modulation modes and signal-to-noise ratios. The dataset contains 24 modulation styles (such as OOK, 4ASK, 8ASK, BPSK, QPSK, etc.). The signal-to-noise ratio ranges from -20 to 30 dB, with an interval of 2dB, and there are 26 signal-to-noise ratios in total (i.e., -20, -18, ..., 28, 30 dB). Each modulation mode has 4096 data at each signal-to-noise ratio, totaling 2555904 data (24 modulation modes x 26 signal-to-noise ratios x 4096 data). The data format is IQ data, and the dimension of each sample is (2, 1024).

[0036] Therefore, in order to collect training data of the source dataset, this embodiment needs to select a total dataset, such as DeepSig RadioML 2018.01A. This dataset needs to have sufficiently rich distribution and semantic information to match the downstream task dataset.

[0037] Based on various influencing factors in wireless communication channels, this embodiment designs a dynamic simulation channel. By introducing the Doppler effect and multipath effect and dynamically changing the parameters, the real-time change of the channel is realized to simulate the influence of the real channel on the signal, thereby enhancing the data set.

[0038] The Doppler effect describes the frequency change caused by the relative motion of the wave source or receiver. In wireless communication, the Doppler shift affects the spectrum of the signal, usually resulting in a frequency offset and phase change. In this embodiment, the parameters Control the influence of Doppler effect on signal frequency deviation. The specific expression is as follows: (4); in, for The signal is processed by the Doppler effect at all times. express The input signal at time, Represents the Doppler shift parameter.

[0039] The multipath effect is modeled using the Rician fading model, which assumes that in addition to being affected by multiple scattering paths (i.e., non-line-of-sight signals), the signal also has a main path (line-of-sight signal), where the main path signal is stronger than the scattering path signal. In this embodiment, the Rician factor is used. Controls the Rice fading model. This parameter represents the ratio of the signal power of the main path to that of the scattered path. A larger K factor means that the main path signal plays a dominant role. The specific expression is as follows: (5); in, represents the Rice factor, which is the ratio of the direct-line path power to the scattered path power. represents the direct-view path signal, is the path gain, represents the scattered path signal, It is The scattering path gain, It is The delay under the scattering path is for The input signal at time, is the index of the scattering path, Is an imaginary unit.

[0040] Combining the above Doppler effect and multipath effect, we can get: (6); (7); Substituting formula (6) and (7) into the Rice fading model of formula (5), we can get Signal data after time enhancement It can be expressed as: (8).

[0041] The specific operations of preprocessing are as follows: First, the IQ two-way signals in the data set are synthesized into a complex signal , and then according to the above Expression, get the enhanced signal data , by updating the channel's Rice factor every several sampling points and Doppler shift parameters , the dynamic change of the channel can be realized, and the change rules of these two parameters can be set by themselves, which can be linear or nonlinear. In this way, multiple enhanced signals of the original signal are generated, which can allow the training modulation style recognition network architecture search framework to obtain richer training data and improve the recognition efficiency and robustness of the modulation style recognition network architecture search framework.

[0042] After the enhanced signal data Finally, in view of the influence of noise on the signal, the Butterworth model is used to design a low-pass filter, which effectively removes the influence of noise while retaining the original data information as much as possible. The frequency response of the Butterworth filter can be expressed as: (9); in, is the transfer function of the filter, represents the independent variable in the frequency domain, represents the cutoff frequency of the low-pass filter, Indicates the order of the filter.

[0043] In this embodiment, the Butterworth filter order used is 4th order, and the normalized cutoff frequency is set to 0.2. The filter frequency response obtained is as follows: Figure 2 shown.

[0044] Enhanced signal The process of passing through the Butterworth filter can be expressed as: (10); in, express The inverse Fourier transform of Represents vector convolution operation, the resulting signal This is the signal after data preprocessing.

[0045] Figure 3 This is an example of a signal denoised by low-pass filtering. The original signal modulation mode is QPSK, and the signal-to-noise ratio is -2. For ease of observation, only 200 points are taken here for plotting. On the left are the constellation diagram (a), amplitude and phase diagram (b), and spectrum diagram (c) of the original signal, and on the right are the constellation diagram (d), amplitude and phase diagram (e), and spectrum diagram (f) generated after low-pass filtering and denoising. It can be clearly seen that the outline of the signal constellation diagram after denoising is more obvious, the amplitude and phase changes are more stable, and the spectrum diagram has a better effect on suppressing high-frequency noise.

[0046] For this dataset, k categories are randomly sampled each time to form a sub-dataset, that is, sampling is performed from the distribution of the task. Then, for each sub-dataset, a modulation style recognition network architecture is randomly sampled in the search space, and the modulation style recognition network architecture is trained using uniform training parameters to obtain the true accuracy of the modulation style recognition network architecture on the sub-dataset, that is, the (sub-dataset, modulation style recognition network architecture, true accuracy) task triplet is obtained as cross-training data.

[0047] Based on the sub-datasets in the cross-scene dataset and the modulation style recognition network architecture, a mean square error loss is constructed to train the generated guidance predictor and performance predictor, thereby obtaining the modulation style recognition network architecture applied to the cross-scene dataset.

[0048] (a5) Based on the sub-datasets in the cross-scene dataset and the modulation style recognition network architecture, a generative guidance predictor is constructed with mean squared error loss training. The generative guidance predictor is used to guide the score network in the latent space to generate the modulation style recognition network architecture through diffusion.

[0049] The role of the generation guide predictor is to predict the performance of the intermediate results in the denoising process of the diffusion model and guide the diffusion model to generate the modulation style recognition network architecture in the latent space. After obtaining the cross-scenario dataset, the generation guide predictor needs to be trained. In the training phase, the guide predictor is used in the task distribution. This allows the predictor to accurately predict the performance of unknown datasets without additional training. To this end, the predictor contains a dataset encoder that encodes and learns relevant features of the dataset.

[0050] The cross-scenario dataset consists of a triplet of (sub-dataset, modulation style recognition network architecture, true accuracy). In the training process of the modulation style recognition network architecture search framework, the input of the guide predictor is generated, which includes three parts: the sub-dataset, the time t, and the network architecture latent vector representation after the architecture passes through the architecture encoder, and then the noise result after the forward diffusion process at time t (i.e., the input of the scoring network). ), the noise adding process refers to Figure 4 The output is the prediction of the true accuracy.

[0051] Generate bootstrap predictor Input-based modulation pattern recognition network architecture and cross-scenario datasets The mean square error loss is constructed to minimize the predicted accuracy and true accuracy s of the model on each task sampled from the cross-scenario dataset distribution. The mean square error loss is as follows: (11); in, Indicates time, The modulation style recognition network architecture is represented by the modulation style recognition network architecture after the architecture encoder, and then after time The result of forward diffusion and noise addition is is the sub-dataset obtained by sampling the source data set, is the prediction accuracy, Identifying network architectures for modulation patterns exist The true accuracy of For parameters Regarding the task triplet The loss function is The task triple , For the distribution of tasks, Represents the trainable parameters for generating a bootstrapped predictor.

[0052] During the deployment and execution of the modulation style recognition network architecture search framework, the architecture input for generating the guided predictor is the intermediate result of the diffusion model denoising, the denoised time step, and the target dataset features.

[0053] To make the generative guidance predictor aware of time steps , we need to pass a relevant time embedding to the generative guided predictor, first compute half the embedding dimension, compute the floor of the embedding, compute the exponential form of the embedding, and then compare the result with the time step The embeddings are multiplied together, and finally the sine and cosine embeddings are merged. If the embedding dimension is odd, zero padding is also required. The generative guided predictor itself consists of convolutions, fully connected layers, and activation functions. The temporal embeddings are added after the predictor processes the representation of the neural network, and then processed by multiple layer blocks. The purpose of the generative guided predictor is to add conditional gradients to each denoising step during the denoising generation process of the diffusion model, so that the denoising process of the score network moves in the direction that meets the required conditions.

[0054] In order to make the cross-domain predictor aware of the dataset, a dataset encoder needs to be designed. The effectiveness of the cross-domain predictor depends on the accuracy of the dataset encoder in capturing the distribution of the target dataset and the accuracy of extracting information related to the generator and predictor targets. This embodiment does not directly input the data in the cross-scene dataset, but inputs the feature embedding representation in the cross-scene dataset. For example, its representation can be extracted from a pre-trained ResNet18, using the feature layer before classification prediction. In order to compress the entire instance in the cross-dataset source dataset D into a single potential code z, the set encoder should handle input sets of any size and summarize consistent information agnostic to the order of instances, that is, it has permutation invariance.

[0055] Therefore, the ensemble encoder for generating guided predictors stacks two permutation invariant modules with attention-based learnable parameters. The low-level intra-class encoder captures class prototypes that reflect label information, and the high-level inter-class encoder considers the relationship between class prototypes and aggregates them into a latent vector. The ensemble encoder structure adopted in this embodiment simulates high-order interactions between ensemble elements, enabling the generator and predictor to effectively extract useful information to achieve each goal.

[0056] Therefore, the dataset encoder of this embodiment is designed as follows: for a given cross-dataset source dataset ,in , is the set of instances in C categories of the cross-dataset source dataset, is the set of labels corresponding to the instances of the source dataset across the dataset. First, randomly sample a class c. ,in is a dimensional feature vector, and , for a category that contains The sampled instances are input into the intra-class encoder IntraSetPool, and for each class Encode to get the class prototype The set representation of each class is then fed into the inter-class encoder InterSetPool to generate the dataset representation , as shown below: (12); (13); For the specific architecture of the dataset encoder, the main architecture is based on attention. Fig.10 The architecture of the multi-head attention module and its derived self-attention block are shown. First, define the multi-head attention block (MultiheadAttention Block, MAB): (14); (15); in, is the layer regularization, is a multi-head attention mechanism, is a multi-head attention block, It is a row-wise feedforward layer, which processes each instance of the input independently and identically. is the regularized result of the sum of the input features and their multi-head self-attention transformation. Then, A row-by-row feed-forward layer is made, that is, each sample is processed independently and identically, and then Add.

[0057] Based on the multi-head attention block, the architecture of the dataset encoder adopts the existing Set Attention Block (SAB) and Pooling by Multi-head Attention (PMA). The former SAB uses self-attention to learn the features of each element in the set, while PMA pools the input features into k representative vectors. The Set Attention Block is an attention-based block, such as Fig.11 As shown, it makes the characteristics of all instances in the collection reflect the relationship between itself and other instances: (16); (17); in, is the collective attention block, is a multi-layer perceptron, Based on the multi-head attention mechanism MH(Q,K,V), the query Q, key K, and value V are all set as the input set X.

[0058] The features encoded from the pooled attention blocks can be processed by PMA on a learnable seed vector Merge above, by modifying Calculate to generate k vectors: (18); (19); in, It is multi-head attention pooling.

[0059] where k can be of any size, and by default is set to 1 to generate a single latent vector. In order to extract consistent information that does not depend on the order and size of the input elements, the encoding function should be constructed by stacking permuted equivariant layers E, is an element in the set Z, which satisfies any permutation on the set Z The following conditions: (20); Since all components in the attention block and multi-head attention pooling are row-computation functions, the two are permutation equivalent according to the definition of the equation.

[0060] (a6) A performance predictor trained with mean squared error loss is constructed based on a sub-dataset in the cross-scene dataset and a modulation style recognition network architecture. The performance predictor is used to predict the true performance of the architecture generated by the estimated diffusion model.

[0061] For the performance predictor, its role is to predict the performance of the modulation pattern recognition network architecture generated by the diffusion model. The same dataset encoder as the generated guided predictor is used to encode and learn the relevant features of the dataset.

[0062] During the training process of the modulation style recognition network architecture search framework, the input of the performance predictor consists of two parts, including the sub-dataset and the modulation style recognition network architecture, and the output is the prediction of the true accuracy.

[0063] Performance Predictor Neural network architecture based on input and cross-scenario datasets The mean square error loss is constructed to minimize the predicted accuracy and true accuracy s of the model on each task sampled from the cross-scenario dataset distribution. The mean square error loss is as follows: (twenty one); in, represents the modulation pattern recognition network architecture, is the sub-dataset obtained by sampling the source data set, is the prediction accuracy, Identifying network architectures for modulation patterns exist The true accuracy of For parameters Regarding the task triplet The loss function is The task triple , For the distribution of tasks, Represents the trainable parameters of the performance predictor.

[0064] The performance predictor consists of a bidirectional graph encoder and a dataset encoder, and finally the prediction results are output by the fully connected layer and the activation layer. Among them, the bidirectional graph encoder consists of a forward graph encoder and a reverse graph encoder. Both the forward graph encoder and the reverse graph encoder are directed acyclic graph encoders. The difference between the two is the execution order.

[0065] In the forward graph encoder, for a given neural network architecture ,according to Perform message passing to the nodes from the predecessor in topological order and iteratively update the hidden state : (twenty two); (twenty three); Among them, the UPDATE function is the gated recurrent unit (GRU), It is the nodes, is the time step Incoming The node information, the function AGGREGATE is a gating function consisting of a mapping and a fully connected layer. is connected to the node The set of predecessor nodes of For Node At the moment The hidden state representation of for One-hot encoding of the operation type, For Node At the moment The hidden state representation of For collection The leading node in .

[0066] For the starting node with an empty predecessor set , outputs the zero vector as The hidden state of the end node is used as the output of the forward graph encoder. . In addition, bidirectional encoding can be utilized to reverse the node order to perform the encoding process. In this case, the last node becomes the starting point. The output of the reverse graph encoder is the last hidden state of the starting node Finally, the output of the forward graph encoder and the reverse graph encoder They are connected as the output of the bidirectional graph encoding graph and combined with the encoding result of the dataset encoder and then passed through a fully connected layer to obtain the performance prediction of the neural network.

[0067] Step 2: Guide the search for modulation style recognition network architecture across data sets, and generate a model to generate an optimal modulation style recognition network architecture based on a given modulation style data set; The overall process of generating architectures on the target dataset is as follows: Generate guided predictors and diffusion models. Generate a batch of modulation style recognition network architectures through the conditional diffusion process, and the performance predictor ranks the generated modulation style recognition network architectures. Select the top k candidate architectures, and use existing performance evaluation methods to obtain the performance of the modulation style recognition network architectures. These modulation style recognition network architectures and their corresponding performance are added to the fine-tuning training set. Weight the data according to the performance of the modulation style recognition network architecture, and use these data to fine-tune the generation of guided predictors and performance predictors. Repeat this process until the expected number of architectures with verified performance is reached. Finally, select the architecture with the best performance as the neural network architecture for subsequent training and deployment.

[0068] When the downstream dataset is a target dataset composed of electromagnetic signals transmitted through a channel, the modulation style recognition network architecture search framework is deployed on the target dataset. First, the conditional diffusion model is defined. After the basic diffusion model is trained, the model parameters are fixed, and then guidance is added to the diffusion process. Inspired by the parameterized model guidance scheme, a parameterized generation guidance predictor is added to the generation framework to guide the basic diffusion model to generate an architecture that meets specific goals. is the desired property that the neural network architecture is expected to satisfy (e.g., high accuracy, low time delay, high robustness). Then, the target attribute in the target dataset features is transformed into The information of is contained in the distribution function, from the conditional distribution Generate neural network architecture: (twenty four); According to Bayes' theorem , the conditional score function can be decomposed into the sum of two gradients: (25); The unconditional diffusion model is approximated by the scoring network , an estimate is needed here ,because Representing the neural network architecture Satisfy target attributes The log-likelihood of Fitting: (26); The introduction of the generative guidance predictor is equivalent to the guidance of the score network diffusion, which predicts whether the given modulation style recognition network architecture has the target attribute. , and then add the corresponding gradient to the denoising generation process to generate a modulation style recognition network architecture with different properties: (27); in, is a constant that determines the strength of guidance for the generated guided predictor. Unlike training a normal predictor, the generated guided predictor predicts the time step Moment modulation pattern recognition network architecture The properties of , which is equivalent to the performance of the intermediate result between the modulation style recognition network architecture and noise. By combining with different generation-guided predictors, it is easily applicable to various types of NAS tasks without retraining the basic generation model, making it applicable to various NAS scenarios.

[0069] like Figure 4 As shown in the figure, through the conditional diffusion process, the task can be quickly deployed on the target dataset. For each category of data in the source dataset, n data are randomly sampled for each category, for example, n=20, as the cross-scenario dataset feature input pre-trained feature extractor and dataset encoder, that is, the data of the entire source dataset is not needed for support. Then the diffusion model samples from Gaussian noise and starts the denoising process. In this process, the intermediate neural network architecture and cross-scenario dataset features are input to the generative guide predictor, and the performance of the denoising result is predicted at each step. Then the gradient of the result to the input is added to the denoising process, that is, the generative guide predictor provides conditional information for the denoising process of the diffusion model through a method similar to gradient descent, and optimizes the result of the denoising process in the direction of high performance. The final denoising result is input into the architecture decoder, and the decoding obtains the connection matrix and operation matrix representation of the network. At this time, the result in the matrix is ​​non-discrete. Through the binarization method, the value ≥0.5 is set to 1, and the value <0.5 is set to 0, and 01 is obtained to represent the discrete matrix. Then, according to the definition of the search space, determine whether this discrete matrix meets the requirements of the search space, and exclude the construction of neural network architectures that do not belong to the search space, such as multiple operations per node, no operations, non-compliant connections, etc. A batch of generated compliant neural network architectures and cross-scenario dataset features are input into the performance ranking predictor, and the top ranking architecture is selected based on the ranking results of the performance ranking predictor.

[0070] The above content describes how the generative guide predictor and the performance predictor are combined with the diffusion model to generate a modulation style recognition network architecture according to the target data set. Step one provides a good initial value for the training of the generative guide predictor and the performance predictor. In order to further improve the performance of the obtained architecture, fine-tuning can be performed on the target data set. After each architecture is generated, the k modulation style recognition network architectures with the highest performance are selected according to the ranking results of the performance predictor, and the network performance is evaluated using the existing evaluation method to obtain the performance estimate corresponding to each architecture. Then, these data can be used to fine-tune the two predictors. Since the training samples are limited. In order to improve the perception of the two predictors on the architecture representation space, this embodiment updates the two predictors through weighted retraining, thereby obtaining a sample-efficient search algorithm. The intuition of this method is to focus more on the high-performance potential network. During the training process, the training data is weighted according to the performance of the network, and the generative guide predictor and the performance predictor are fine-tuned on the weighted data distribution. For the fine-tuning training set with known performance , for each architecture Assign a weight , indicating the relative likelihood of a sample being sampled during training, and weighting the corresponding loss function during the training phase ,in It is architecture As input, we use the loss function. For the weights, we use the rank-based weights: (28); (29); in is the modulation pattern recognition network architecture, It is architecture The relative weight of the assignment, is the number of architectures in the current fine-tuning training set, is a constant term, It is architecture The evaluation method, is the fine-tuning training set, express Performance over architecture The number of architectures, i.e. architecture Rank in the current training set. By focusing the cross-domain predictor on high-performance samples, a high-performance subspace can be learned using a small amount of data.

[0071] Step 3: Perform parameter training on the optimal modulation pattern recognition network architecture according to the given modulation recognition data set to obtain the modulation pattern recognition network model.

[0072] Take the DeepSig RadioML 2016.10A dataset as an example. This dataset is a dataset for modulation recognition in wireless communications. It contains up to 11 modulation types (such as BPSK, QPSK, 16-QAM, etc.) and covers different noise conditions and channel environments. Each sample is a complex sequence of length 128 with annotated modulation type labels. The dataset is suitable for tasks such as signal classification, modulation recognition, and wireless spectrum sensing using deep learning and traditional machine learning algorithms.

[0073] The data set is randomly divided into training set, validation set and test set in a ratio of 6:2:2. The data preprocessing process refers to the data set preprocessing method for wireless communication modulation identification in step 2. The training set is processed by dynamic simulation channel and low-pass filtering, and the validation set is only processed by low-pass filtering.

[0074] For the modulation style recognition network architecture searched in the training phase, the input is the preprocessed training set, and the output is the classification prediction result of the architecture in the training set.

[0075] The entire training set is recorded as: (30); in, Subscript the sample. is the total number of samples; For the samples, by The IQ of each sampling point is composed of two paths; For the There are a total of class labels for samples categories, using Represents category labels The corresponding one-hot vector is used for subsequent training.

[0076] The optimal modulation pattern recognition network architecture obtained in the above steps is , input sample Identify network architecture via optimal modulation pattern The prediction results are recorded as: (31); in, yes The probability distribution of categories. The training loss function uses cross entropy loss, expressed as: (32); in, Represents the optimal neural network architecture Parameters, and then use the gradient descent method to optimize the parameters , expressed as: (33); in, , Respectively and The model parameters of the optimal modulation style recognition network architecture after iterations, Represents the learning rate of the optimal modulation pattern recognition network architecture training. The above process is iterated repeatedly until the performance of the optimal modulation pattern recognition network architecture on the validation set converges, and a modulation pattern recognition model with good performance can be obtained.

[0077] Step 4: Deploy the modulation pattern recognition model to the channel transmission, and perform modulation pattern recognition on the electromagnetic signal transmitted through the channel.

[0078] For a new modulation style dataset , Samples in It is composed of two channels of IQ data of N sampling points. This embodiment needs to Each sample in The modulation pattern is identified.

[0079] In this embodiment, the test set generated in step 3 is taken as an example as the input of the optimal modulation pattern recognition model. After the input signal is denoised by the low-pass filter in step 2, it is input into the modulation pattern recognition model obtained in step 3, and the obtained prediction result is the modulation pattern recognition result of the signal.

[0080] In this embodiment, when the modulation style recognition network architecture search framework is trained and deployed with an unknown target data set, the modulation style recognition network architecture search framework does not need to be retrained, but can directly use the target data set features to generate the network architecture, which greatly reduces the search overhead and reduces the dependence on the amount of data in the target data set features.

[0081] The key point of this embodiment is the generation of a scoring network based on latent space and a cross-domain predictor for cross-scene datasets (including the generation of guided predictors and performance predictors). The scoring network can be combined with predictors of various task types (accuracy, time delay) without retraining, and the modulation style recognition network architecture is encoded into the latent space through the architecture encoder and decoder, which reduces the difficulty of learning the distribution of the modulation style recognition network architecture. The distribution of the search space is modeled by combining the scoring network with the cross-domain predictor. The predictor for cross-scene datasets only needs to be trained once to be deployed in different datasets, in which the introduction of the dataset encoder plays a key role. The cross-domain predictor learns the performance of the architecture on specific tasks through the dataset features, thereby realizing the prediction and generation of cross-scene datasets. The advantage of this embodiment is that other search methods need to be retrained when deploying searches on various downstream data sets, while this embodiment only needs to pre-train the architecture codec, diffusion model, and cross-domain predictor of cross-scenario data sets once, and can be directly deployed on various downstream data sets, without generating additional search overhead, greatly accelerating the search efficiency, and generating a modulation style recognition network architecture with excellent performance. Moreover, this embodiment does not need to use the entire downstream data set for training during the deployment of the downstream data set, which reduces data dependency and can be applied to scenarios with little data. In addition, this embodiment is scalable, and only needs to adjust the prediction target of the performance predictor based on the cross-scenario data set to generate a modulation style recognition network architecture that meets different properties, including high accuracy, low time delay, high robustness, hardware indicators, etc., to search the network for specific tasks and realize a dedicated network.

[0082] The present invention provides a cross-dataset generation framework, in which the specific network architecture can be modified, including but not limited to the selection of search space, the architectural design of the architecture codec, the architectural design of the diffusion model, the architectural design of the generation guidance predictor and the performance predictor, the usage method and architectural design of the dataset encoder, and for specific tasks and datasets, a suitable deep learning architecture can be selected as a replacement, which has good scalability.

[0083] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A modulation pattern recognition method based on generative neural network architecture search, characterized in that: include: Train a modulation style recognition network architecture search framework to obtain a generative model, a generative guidance predictor, and a performance predictor; Guide the search for modulation style recognition network architecture across data sets, and generate a model to generate the optimal modulation style recognition network architecture based on a given modulation style data set; According to the given modulation recognition data set, the parameters of the optimal modulation pattern recognition network architecture are trained to obtain the modulation pattern recognition network model; The modulation pattern recognition model is deployed on the channel transmission to perform modulation pattern recognition on the electromagnetic signal transmitted through the channel.

2. The modulation pattern recognition method based on generative neural network architecture search according to claim 1 is characterized in that: The training process of the modulation style recognition network architecture search framework is as follows: A parameterized search space of a modulation style recognition network architecture is constructed based on the modulation style recognition task, wherein each modulation style recognition network architecture in the parameterized search space is regarded as a directed acyclic graph, wherein the nodes of the graph represent operations, and the directed edges represent the direction of information transmission, and each style recognition network architecture is represented as an adjacency matrix and an operator operation type matrix; The network architecture encoder-decoder is trained in the parameterized search space, and the modulation style recognition network architecture is encoded into the latent space through the network architecture encoder-decoder; A diffusion model is trained in latent space, and the reverse diffusion process of the diffusion model is used as a generative model of the modulation style recognition network architecture, a performance predictor of the modulation style recognition network architecture, and a generation guidance predictor of the modulation style recognition network architecture.

3. The modulation pattern recognition method based on generative neural network architecture search according to claim 2 is characterized in that: The network architecture encoder-decoder is trained in the parameterized search space of the modulation style recognition network architecture as follows: Noise is added to the modulation style recognition network architecture representation matrix to train the architecture encoder to encode the modulation style recognition network architecture into a potential representation vector, which is input into the architecture decoder after adding noise, and a reconstruction loss is constructed based on the output modulation style recognition network architecture and the input modulation style recognition network architecture to train the architecture encoder-decoder.

4. The modulation pattern recognition method based on generative neural network architecture search according to claim 2, characterized in that: The diffusion model is trained in the latent space as follows: A stochastic differential equation is defined to describe the unconditional forward diffusion process of the modulation style recognition network architecture. The modulation style recognition network architecture in the parameterized search space is gradually noised through the forward diffusion process, and the distribution of the modulation style recognition network architecture is mapped to the known prior distribution to obtain the hidden representation of the architecture after noise addition. An inverse stochastic differential equation is defined to describe the unconditional diffusion process of the modulation style recognition network architecture, and the hidden representation of the noisy architecture is denoised through the reverse diffusion process; In back-diffusion, the trainable parameters of the scoring network are trained to approximate the unknown term score function in the back-diffusion process, thereby obtaining a trained generative model.

5. The modulation pattern recognition method based on generative neural network architecture search according to claim 2, characterized in that: 5-1) Train the generative guidance predictor of the modulation style recognition network architecture in the latent space, specifically: Build a cross-scenario dataset, including sub-datasets, modulation style recognition network architecture, and true accuracy triples; Input the modulation style recognition network architecture representation matrix into the architecture encoder for encoding to obtain the latent space representation vector of the architecture; According to the forward diffusion process, choose the time , diffuse and add noise to the latent space vector; The sub-dataset and the latent space vector after noise are input into the guided predictor to predict the corresponding true accuracy. The mean square error loss training is constructed based on the predicted accuracy and the true accuracy to generate the guided predictor model. During the deployment and execution of the modulation style recognition model, the architecture inputs of the generated guided predictor are the intermediate results of the diffusion model denoising, the denoised time steps, and the target dataset features; 5-2) Train the performance predictor of the modulation style recognition network architecture, specifically: Build a cross-scenario dataset, including sub-datasets, modulation style recognition network architecture, and true accuracy triples; Input the sub-dataset and the modulation style recognition network architecture representation matrix into the performance predictor, predict the corresponding true accuracy, and build a mean square error loss training performance predictor model based on the predicted accuracy and true accuracy; During the deployment and execution of the modulation style recognition model, the input of the performance predictor is the modulation style recognition network architecture finally generated by the diffusion model denoising and the characteristics of the target dataset.

6. The modulation pattern recognition method based on generative neural network architecture search according to claim 4 is characterized in that: The optimization expression for the trainable parameters in the generative model is as follows: in, are the trainable parameters after the scoring network is trained, are the trainable parameters of the scoring network, represents the time of the diffusion process, Indicates expectation, Indicates time-based Given a positive weighting function, is the 0th moment, i.e. the function mapping of the initial modulation style recognition network architecture, For the moment Function mapping of the modulation pattern recognition network architecture, Indicates time 0 expectations, Indicates that in a given In case of moment hour expectations, Indicates and time is the input score network, is the unknown term score function, About The gradient of Based on time of The marginal probability distribution of .

7. The modulation pattern recognition method based on generative neural network architecture search according to claim 5, characterized in that: The cross-scene dataset is constructed as follows: Preprocess the data set used for wireless communication modulation recognition to obtain the source data set, randomly sample k categories from the source data set each time to form a sub-data set, randomly sample a modulation style recognition network architecture in the parameterized search space, train the modulation style recognition network architecture to obtain the true accuracy on the sub-data set, and use the task triple consisting of the sub-data set, modulation style recognition network architecture, and true accuracy as cross-training data to construct a cross-scenario data set; A mean square error loss is constructed based on the sub-datasets and modulation style recognition network architecture in the cross-scene dataset to train a cross-domain predictor to search the modulation style recognition network architecture of the source dataset. The generative model generates the optimal modulation style recognition network architecture according to the task dataset. The cross-domain predictor includes a generation guidance predictor and a performance predictor.

8. The modulation pattern recognition method based on generative neural network architecture search according to claim 7 is characterized in that: The construction process of the source dataset is as follows: Dynamic simulation channel: Introduce Doppler effect and multipath effect to simulate the impact of real channel on signal and construct dynamic simulation channel: Data enhancement: The IQ signals in the data set are synthesized into an input complex signal. After the input complex signal passes through the dynamic simulation channel, the enhanced signal data is obtained. The K factor and Doppler frequency shift parameters of the channel are updated every several sampling points to realize the dynamic change of the channel. Filtering: The enhanced signal data is filtered based on a low-pass filter, and the processed data is used as the source data set; The expression of dynamic simulation channel is as follows: in, for Signal data after time enhancement, is the Rice factor, is the Doppler frequency shift parameter, for The input signal at time, for The input signal at time, is the path gain, is the number of scattering paths, It is The scattering path gain, It is The delay under the scattering path is Is an imaginary unit.

9. The modulation pattern recognition method based on generative neural network architecture search according to claim 5, characterized in that: The generation guide predictor and the performance predictor are both provided with a data set encoder, and the structures of the two data set encoders are consistent; For generating a guided predictor, the mean square error loss constructed by the predicted accuracy and the true accuracy is as follows: in, Indicates time, The modulation style recognition network architecture is represented by the modulation style recognition network architecture after the architecture encoder, and then after time The result of forward diffusion and noise addition is is the sub-dataset obtained by sampling the source data set, For prediction accuracy, Identifying network architectures for modulation patterns exist The true accuracy of For parameters Regarding the task triplet The loss function is The task triple , For the distribution of tasks, represents the trainable parameters for generating the guided predictor; For the performance predictor, the mean square error loss constructed by the predicted accuracy and the true accuracy is as follows: in, For prediction accuracy, For parameters Regarding the task triplet The loss function is Representation performance predictor The trainable parameters of .

10. The modulation pattern recognition method based on generative neural network architecture search according to claim 5, characterized in that: A guided predictor and a performance predictor are generated by training a cross-scenario dataset. The generated guided predictor is used to guide the score network to generate a latent space representation of the modulation style recognition network architecture through a diffusion model in the latent space. The architecture decoder maps the latent space representation generated by the diffusion model to the modulation style recognition network architecture in the original search space. The performance predictor is used to predict the performance of the mapped modulation style recognition network architecture. During the deployment and execution of the modulation style recognition network architecture search framework, a generation guided predictor guides the generation model to generate a set of modulation style recognition network architectures applied to a given modulation recognition data set, the modulation style recognition network architecture set is sorted in descending order based on the performance predictor, and the modulation style recognition network architecture with the highest performance ranking is selected for performance evaluation. The selected modulation style recognition network architecture and the corresponding evaluation accuracy are used to fine-tune the generation guided predictor and the performance predictor to adapt to the generation of a given modulation recognition data set.

Images