A Modulation Pattern Recognition Method Based on Generative Neural Network Architecture Search

Abstract

This invention discloses a modulation pattern recognition method based on generative neural network architecture search, relating to the fields of artificial intelligence and wireless communication. The method includes: training a modulation pattern recognition network architecture search framework to obtain a generative model, a generative guidance predictor, and a performance predictor; guiding a cross-dataset modulation pattern recognition network architecture search, whereby the generative model generates the optimal modulation pattern recognition network architecture based on a given modulation pattern dataset; training the parameters of the optimal modulation pattern recognition network architecture based on a given modulation recognition dataset to obtain a modulation pattern recognition network model; and deploying the modulation pattern recognition model onto channel transmission to perform modulation pattern recognition on electromagnetic signals transmitted through the channel. This modulation pattern recognition method can quickly and automatically search for the optimal modulation pattern recognition network model based on a given modulation pattern dataset, thus rapidly adapting to different application scenarios. Simultaneously, this method can significantly reduce search overhead and lower the requirement for large datasets during the network architecture search process.

Description

Technical Field

[0001] This 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 Technology

[0002] In modern communication systems, signal modulation pattern recognition technology is a crucial component for communication quality assurance and spectrum resource management. Different types of signal modulation patterns are widely used in wireless transmission to meet diverse 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), and QPSK (quadrature phase modulation).

[0003] With increasing signal complexity, automatic modulation type identification has become a critical technical problem that urgently needs to be solved. In wireless communication, especially in the context of spectrum sharing and interference management, timely and accurate identification of signal modulation patterns is crucial for improving the system's anti-interference capability and optimizing spectrum resource utilization. Traditional modulation identification methods rely on pre-set feature extraction and classification algorithms, which have achieved certain application results. However, with the diversity of signals and changes in noise environments, the identification accuracy and robustness of traditional methods face challenges.

[0004] With the continuous development of machine learning and deep learning technologies, data-driven modulation recognition methods have gradually become a research hotspot. However, current methods typically require manual design and adjustment of the neural network architecture, and the experience in designing model architectures often comes from other fields such as image processing and vision. This means that the network model may not be well adapted to the modulation recognition problem of communication signals. Furthermore, current recognition methods still exhibit poor performance under low signal-to-noise ratio conditions. Therefore, how to automatically generate a reliable model architecture suitable for communication modulation recognition is a very promising research direction.

[0005] Neural Architecture Search (NACS) aims to automate the network architecture design process, discovering efficient neural network structures for specific problems with minimal human intervention. Early NACS techniques primarily relied on reinforcement learning and evolutionary algorithms, updating the search strategy by evaluating the found network architectures. This approach consumed enormous computational resources and was unsuitable for large search spaces and datasets. Differentiable architecture search accelerated the search, but the network architectures selected from the search space often fell into suboptimal states. Therefore, techniques have been developed to model the search space using generative models, directly generating architectures from promising regions. However, these methods share a common problem: the search algorithm needs to be retrained on each dataset, resulting in significant resource waste when searching across multiple datasets. Summary of the Invention

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

[0007] This invention proposes a modulation pattern recognition method based on generative neural network architecture search; comprising:

[0008] A modulation pattern recognition network architecture search framework is trained to obtain a generative model, a generative guidance predictor, and a performance predictor;

[0009] Guide the search for modulation pattern recognition network architecture across datasets, and generate the optimal modulation pattern recognition network architecture based on the given modulation pattern dataset.

[0010] The optimal modulation pattern recognition network architecture is trained with parameters based on the given modulation recognition dataset to obtain the modulation pattern recognition network model.

[0011] The modulation pattern recognition model is deployed on the channel transmission to perform modulation pattern recognition on the electromagnetic signals transmitted through the channel.

[0012] Furthermore, the training process of the modulation pattern recognition network architecture search framework is as follows:

[0013] A parameterized search space for modulation pattern recognition network architecture is constructed based on the modulation pattern recognition task. Each modulation pattern recognition network architecture in the parameterized search space is regarded as a directed acyclic graph. The nodes of the graph represent operations, and the directed edges represent the direction of information transmission. Each pattern recognition network architecture is represented as an adjacency matrix and an operator operation type matrix.

[0014] Train the network architecture encoder-decoder in the parameterized search space, and encode the modulation pattern recognition network architecture into the latent space using the network architecture encoder-decoder;

[0015] The diffusion model is trained in the latent space, and the back-diffusion process of the diffusion model is used as the generative model of the modulation pattern recognition network architecture, the performance predictor of the modulation pattern recognition network architecture, and the generation guidance predictor of the modulation pattern recognition network architecture.

[0016] Furthermore, the network architecture encoder and decoder are trained in the parameterized search space of the modulation pattern recognition network architecture, as follows:

[0017] Noise is added to the modulation pattern recognition network architecture representation matrix to train the architecture encoder to encode the modulation pattern recognition network architecture into a latent representation vector. Noise is added to the latent representation vector and then input into the architecture decoder. A reconstruction loss is constructed based on the output modulation pattern recognition network architecture and the input modulation pattern recognition network architecture to train the architecture encoder and decoder.

[0018] Furthermore, the diffusion model is trained in the latent space, specifically as follows:

[0019] A stochastic differential equation is defined to describe the unconditional forward diffusion process of the modulation pattern recognition network architecture. The modulation pattern recognition network architecture in the parameterized search space is gradually noise-added through the forward diffusion process, and the distribution of the modulation pattern recognition network architecture is mapped to the known prior distribution to obtain the noisy architecture hidden representation.

[0020] We define an inverse stochastic differential equation to describe the unconditional diffusion process of the modulation pattern recognition network architecture, and denoise the hidden representation of the noisy architecture through the reverse diffusion process.

[0021] In backdiffusion, the trainable parameters of the scoring network are trained to approximate the score function of the unknown terms in the backdiffusion process, thereby obtaining the trained generative model.

[0022] Further, 5-1) the generative guided predictor of the modulation pattern recognition network architecture is trained in the latent space, specifically as follows:

[0023] Construct a cross-scenario dataset, including a subset of datasets, a modulation pattern recognition network architecture, and true accuracy triples;

[0024] The modulation pattern recognition network architecture representation matrix is ​​input into the architecture encoder for encoding to obtain the latent space representation vector of the architecture;

[0025] Based on the forward diffusion process, select the time. , to diffuse and add noise to the latent space vector;

[0026] The subset of data and the noisy latent space vectors are input to generate a guided predictor, which predicts the corresponding true accuracy. Based on the prediction accuracy and the true accuracy, the mean squared error loss is constructed to train and generate the guided predictor model.

[0027] During the deployment and execution of the modulation pattern recognition model, the architectural inputs for generating the guiding predictor are the intermediate results of the diffusion model denoising, the denoising time steps, and the features of the target dataset.

[0028] 5-2) Training the performance predictor of the modulation pattern recognition network architecture, specifically:

[0029] Construct a cross-scenario dataset, including a subset of datasets, a modulation pattern recognition network architecture, and true accuracy triples;

[0030] Input the performance predictor with the subset of data and the modulation pattern recognition network architecture representation matrix to predict the corresponding true accuracy. Based on the predicted accuracy and the true accuracy, construct the mean squared error loss to train the performance predictor model.

[0031] During the deployment and execution of the modulation pattern recognition model, the input to the performance predictor is the modulation pattern recognition network architecture finally generated by the diffusion model denoising and the features of the target dataset.

[0032] Furthermore, the optimized representation of the trainable parameters in the generative model is as follows:

[0033]

[0034] in, These are the trainable parameters after training the scoring network. These are the trainable parameters of the scoring network. Indicates the moment of the diffusion process. Indicates taking the expected value. Indicates time-based Given a positive weighting function, At time 0, i.e., the function mapping of the initial modulation pattern recognition network architecture, For a moment Function mapping of time-modulation pattern recognition network architecture Indicates time 0 Expectations Indicates that in a given In the case of time hour Expectations Indicates and time For the input scoring network, For a fractional function with unknown terms, For about gradient, For time-based of The marginal probability distribution.

[0035] Furthermore, the cross-scenario dataset is constructed as follows:

[0036] The dataset for wireless communication modulation recognition is preprocessed to obtain the source dataset. Each time, k categories are randomly sampled from the source dataset to form a subset. A modulation pattern recognition network architecture is randomly sampled in the parameterized search space. The true accuracy of the modulation pattern recognition network architecture on the subset is trained. The task triple consisting of the subset, the modulation pattern recognition network architecture, and the true accuracy is used as cross-training data to construct a cross-scene dataset.

[0037] The mean squared error loss is constructed based on a subset of the cross-scene dataset and the modulation pattern recognition network architecture to train a cross-domain predictor to search for the modulation pattern recognition network architecture of the source dataset. The generation model generates the optimal modulation pattern recognition network architecture based on the task dataset. The cross-domain predictor includes a generation guidance predictor and a performance predictor.

[0038] Furthermore, the process of constructing the source dataset is as follows:

[0039] Dynamic Simulated Channel: The Doppler and multipath effects are introduced to simulate the impact of real channels on signals, thus constructing a dynamic simulated channel.

[0040] Data augmentation: The I and Q signals in the dataset are combined into an input complex signal. After the input complex signal passes through a dynamic simulation channel, the enhanced signal data is obtained. By updating the K factor and Doppler frequency shift parameters of the channel at several sampling points, the dynamic change of the channel is realized.

[0041] Filtering: The enhanced signal data is filtered using a low-pass filter, and the processed data is used as the source dataset.

[0042] The expression for the dynamic simulated channel is as follows:

[0043]

[0044] in, for The signal data after real-time enhancement Rice factor, For Doppler frequency shift parameters, for Input signal at time, for Input signal at time, For path gain, The number of scattering paths, It is Each scattering path gain, It is Delay under each scattering path It is the imaginary unit.

[0045] Furthermore, both the generation guidance predictor and the performance predictor are equipped with a dataset encoder, and the two dataset encoders have the same structure.

[0046] For the generative guided predictor, the mean squared error loss constructed from the prediction accuracy and the true accuracy is as follows:

[0047]

[0048] in, Indicates time, The modulation pattern recognition network architecture is represented by the architecture encoder, and then after time... The result after forward diffusion with added noise, This is a subset of the sampled source dataset. For prediction accuracy, Modulation pattern recognition network architecture exist The true accuracy on For parameters Regarding the task triplet in the following situations loss function, For the task triple , For the distribution of tasks, This represents the trainable parameters for generating the guided predictor;

[0049] For the performance predictor, the mean squared error loss constructed from the prediction accuracy and the true accuracy is as follows:

[0050]

[0051] in, For prediction accuracy, For parameters Regarding the task triplet in the following situations loss function, Indicates performance predictor Trainable parameters.

[0052] Furthermore, a guided predictor and a performance predictor are trained using a cross-scene dataset. The guided predictor is used to guide the scoring network to generate the latent space representation of the modulation pattern recognition network architecture through the diffusion model in the latent space. The architecture decoder maps the latent space representation generated by the diffusion model to the modulation pattern recognition network architecture in the original search space. The performance predictor is used to predict the performance of the mapped modulation pattern recognition network architecture.

[0053] During the deployment and execution of the modulation pattern recognition network architecture search framework, the generative model is guided by the generation guide predictor to generate a set of modulation pattern recognition network architectures applicable to a given modulation recognition dataset. The set of modulation pattern recognition network architectures is sorted in descending order based on the performance predictor. The modulation pattern recognition network architectures with the highest performance ranking are selected for performance evaluation. The generation guide predictor and the performance predictor are fine-tuned using the selected modulation pattern recognition network architectures and their corresponding evaluation accuracy to adapt to the generation of the given modulation recognition dataset.

[0054] The advantages of the modulation pattern recognition method based on generative neural network architecture search provided by this invention are as follows: Based on neural network architecture search technology, it automatically searches for the most suitable network architecture for the modulation pattern recognition task from the dataset, avoiding manual intervention in traditional methods, thereby improving the efficiency and accuracy of the recognition process. Simultaneously, to overcome the need for existing network architecture search algorithms to be re-run on each dataset, a predictor-guided diffusion model generation method is used. The neural network architecture is encoded in the latent space, and a pre-trained predictor guides the diffusion model to generate neural network architectures across different scene datasets. When the modulation pattern recognition network architecture search framework is trained and deployed to an unknown dataset, the framework does not need to be retrained; instead, it can directly utilize the features of the target dataset to generate the network architecture, greatly reducing search overhead and decreasing dependence on the amount of data in the target dataset features. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the present invention;

[0056] Figure 2 This is a spectrum diagram of a Butterworth low-pass filter;

[0057] Figure 3 It includes the constellation diagram, amplitude and phase diagram, and spectrum diagram of the original signal and the denoised signal;

[0058] Figure 4 This is a flowchart of the process of generating an excellent neural network architecture;

[0059] Figure 5 This is a schematic diagram of the architecture of the codec in the latent space;

[0060] Figure 6 This is a schematic diagram of the forward and reverse diffusion processes in a diffusion model;

[0061] Figure 7 This is a schematic diagram illustrating the training and deployment of predictors across different datasets.

[0062] Figure 8 This is a schematic diagram of a certain neural network architecture in NAS-Bench-201, along with its corresponding adjacency matrix representation and node operation matrix representation.

[0063] Figure 9 This is a schematic diagram of the search space in NAS-Bench-201.

[0064] Figure 10 This is a schematic diagram of the multi-head attention module structure;

[0065] Figure 11 This is a schematic diagram of the attention block structure. Detailed Implementation

[0066] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the 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 modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0067] like Figures 1 to 11 As shown, this invention proposes a modulation pattern recognition method based on generative neural network architecture search. By training a predictor across datasets, it achieves rapid deployment on the target dataset, enabling the rapid search for excellent neural network architectures for unknown tasks. The target dataset consists of electromagnetic signals transmitted through a channel. Specifically, it includes the following steps one to four:

[0068] Step 1: Train the modulation pattern recognition network architecture search framework to obtain the modulation pattern recognition network architecture generation model, the modulation pattern recognition network architecture performance predictor, and the modulation pattern recognition network architecture generation guided predictor.

[0069] The training process of the modulation pattern recognition network architecture search framework is as follows (a1) to (a7):

[0070] The details are as follows:

[0071] (a1) Constructing a parameterized search space for a modulation pattern recognition network architecture based on the modulation pattern recognition task;

[0072] The parameterized search space is the 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 nodes represent operations and directed edges represent the direction of information transmission. A network structure with N nodes and F predefined operations is given. , by its operator type matrix and the upper triangular adjacency matrix Defined as A neural network architecture can be represented by an adjacency matrix and by an operation matrix. Thus, each neural network architecture in the parameterized search space can be represented by two matrices. For example... Figure 8 This is a schematic diagram of a neural network architecture in NAS-Bench-201, along with its corresponding adjacency matrix and node operation matrix representations. Within the originally defined search space... Figure 9 In this model, edges represent operations and nodes represent computational flow directions. The two can be converted into each other. That is, each pattern recognition network architecture can be represented as two matrices: an adjacency matrix and an operator operation type matrix.

[0073] (a2) Train the network architecture encoder-decoder in the parameterized search space and encode the modulation pattern recognition network architecture into the latent space;

[0074] Once the representation of the modulation pattern recognition network architecture is determined, the problem of generating the modulation pattern recognition network architecture is transformed into the problem of generating two matrices. Only the corresponding matrices need to be generated to obtain the corresponding modulation pattern recognition network architecture. However, considering that both the adjacency matrix and the operator operation type matrix are 0-1 discrete matrices, the Hamming distance between networks is at least 1, which is detrimental to the diffusion model learning an effective network encoding distribution. For example... Figure 5 As shown, based on the latent diffusion model technique, an architecture encoder-decoder is used to encode the modulation pattern recognition network architecture representation matrix into the latent space, and the latent space representation is used to achieve efficient and high-quality modulation pattern recognition network architecture generation. First, the architecture encoder-decoder encodes the modulation pattern 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 computational resource requirements, allowing the diffusion model to better model the search space, while improving the quality of architecture generation.

[0075] The input to the architecture encoder is the representation matrix of the modulation pattern recognition network architecture, namely the adjacency matrix and the operator operation type matrix. The output is the corresponding encoded latent vector in the latent space. The input to the architecture decoder is the encoded latent vector in the latent space, and the output is the reconstructed prediction of the representation matrix of the modulation pattern recognition network architecture corresponding to the latent vector. During the training phase, the encoder and decoder are connected together for joint training.

[0076] Taking a noise-based encoder-decoder network as an example, its architecture is mainly composed of stacked blocks of convolution, LeakyReLU activation, and InstanceNorm1d normalization. The encoder and decoder each have four blocks. The representation matrix of the modulation pattern 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 encoder to the input. Gaussian noise with a standard deviation of 0.1 is added to the intermediate encoding latent space to achieve distance differentiation between different architectures. Reconstruction loss is used to train the encoder-decoder.

[0077] (a3) Train the diffusion model in the latent space and use the back diffusion process of the diffusion model as the generative model of the modulation pattern recognition network architecture, the performance predictor of the modulation pattern recognition network architecture, and the generative guidance predictor of the modulation pattern recognition network architecture.

[0078] First, a stochastic differential equation (SDE) is defined to describe the unconditional diffusion process of the modulation pattern recognition network architecture. This process describes the transformation from the initial distribution of the modulation pattern recognition network architecture in the search space to a prior noise distribution by gradually adding noise perturbations. Then, the perturbation process can be learned to reverse the process, sampling the structure of the search space from the noise. The forward process maps the modulation pattern recognition network architecture distribution P(A0) to the known prior distribution P(A... T The forward diffusion process of ) is defined as follows:

[0079] (1);

[0080] in, It is the drift factor. It is the diffusion factor. Indicates cross-domain predictor The trainable parameters are those of a standard Wiener process. It is infinitesimal white noise. For a moment The function mapping of the time-space neural network architecture. The solution of this SDE is a continuous set of random variables that track the random trajectory of time t from 0 to T. Subsequently, during the forward diffusion process, the architecture is perturbed with Gaussian noise at each step.

[0081] Each stochastic differential equation has a corresponding inverse stochastic differential equation. In order to sample from the distribution, the reverse time diffusion process corresponding to the forward process is modeled by the following SDE:

[0082] (2);

[0083] in, Indicates time-based of Marginal probability distribution, This represents an infinitesimal negative time step. It is a reverse-time standard Wiener process. To The gradient.

[0084] In order to use the inverse process as a generative model, a scoring network needs to be trained. Trainable parameters of (score-net) fractional functions for unknown terms in the reverse process To approximate, where Given a positive weighting function:

[0085] The optimized representation of the trainable parameters in the scoring network is as follows:

[0086] (3);

[0087] in, These are the trainable parameters after training the scoring network. Indicates taking the expected value. At time 0, i.e., the function mapping of the initial modulation pattern recognition network architecture, For a moment Function mapping of time-modulation pattern recognition network architecture Indicates time 0 Expectations Indicates that in a given In the case of time hour Expectations Indicates and time For the input scoring network, For about gradient, For time-based of The marginal probability distribution.

[0088] The input to the scoring network is the latent vector encoded by the encoder from the representation matrix of the modulation pattern recognition network architecture, and then processed by time... The perturbation results obtained from the forward diffusion perturbation, and the time of the hidden vector perturbation. The output is The prediction fit is such that, since Gaussian noise is added during forward diffusion, the expected mean of the forward diffusion result is the same as the input. The form can be mathematically derived based on the selection of the diffusion process model. Taking the variance-explosive stochastic differential equation as an example, the upper triangular part of the random standard Gaussian noise is set to -1, and then this matrix is ​​added to its transpose to obtain the matrix. , The standard deviation of the diffusion process at time step is std, and the perturbation result of the input to the scoring network is... It is a matrix Multiply by the standard deviation (std), then add the encoded latent vector. The loss function value is the score output of the network. Multiply The standard deviation (std) of the diffusion process at time step is then added to the result. Then, the square is taken to obtain the final loss value.

[0089] Understandably, the diffusion model is a mathematical model, while the scoring network is the network used to fit the unknowns in the diffusion model's formula. The training set of the scoring network consists of all modulation pattern recognition network architectures in the search space after being encoded by the architecture encoder, and the input is after random time steps. (i.e., time) The noisy neural network architecture representation The target output is This embodiment aims to enable the scoring network to remove noise from the noisy representation. Once the scoring network is trained, the inverse process can be used to generate a representation that follows the original distribution. A modulation pattern recognition network architecture. It can sample noise from a known prior distribution (Gaussian distribution) and use... By fitting the unknown terms in the stochastic differential equation of the reverse process, and simulating the reverse process step by step to remove noise, the network representation vector is obtained.

[0090] For the architecture of the diffusion model, a diffusion process based on variance explosion SDE can be adopted. Taking a transform modulation pattern recognition network architecture with cross-attention computational perception as an example as the scoring network, the scoring network has 12 layers, with 8 attention heads in each layer. The dimensions of all feedforward layers are... The size of the hidden layer is Configure SDE , Discrete time step .

[0091] (a4) Preprocess the dataset used for wireless communication modulation recognition to obtain the source dataset. Construct a cross-scene dataset based on the preprocessed dataset. Randomly sample k categories from the source dataset each time to form a subset. Randomly sample a modulation pattern recognition network architecture in the parameterized search space. Train the network architecture to obtain the true accuracy on the subset. Use the subset, the modulation pattern recognition network architecture, and the true accuracy as the task triplet to construct the cross-scene dataset.

[0092] This paper uses the DeepSig RadioML 2018.01A dataset as an example to introduce the preprocessing method for the dataset. This dataset is widely used in automatic modulation recognition research and contains data on various modulation schemes and signal-to-noise ratios (SNRs). The dataset contains 24 modulation schemes (such as OOK, 4ASK, 8ASK, BPSK, QPSK, etc.). The SNR ranges from -20 to 30 dB, with 2 dB intervals, resulting in 26 SNRs (i.e., -20, -18, ..., 28, 30 dB). Each modulation scheme has 4096 data points at each SNR, totaling 2,555,904 data points (24 modulation schemes x 26 SNRs x 4096 data points). The data format is IQ data, and each sample has a dimension of (2, 1024).

[0093] Therefore, in order to collect training data from 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.

[0094] Based on the various influencing factors in wireless communication channels, this embodiment designs a dynamic simulated channel. By introducing the Doppler effect and multipath effect and dynamically changing the parameters, the channel changes in real time to simulate the impact of the real channel on the signal, thereby enhancing the dataset.

[0095] The Doppler effect describes the frequency shift caused by the relative motion of a wave source or receiver. In wireless communication, the Doppler shift affects the signal's spectrum, typically resulting in frequency offset and phase change. In this embodiment, parameters are used... The effect of the Doppler effect on signal frequency shift is controlled. The specific expression is as follows:

[0096] (4);

[0097] in, for The signal after being processed by the Doppler effect at any moment, express Input signal at time, This represents the Doppler frequency shift parameter.

[0098] The multipath effect is modeled using the Rician fading model. The Rician fading model assumes that in addition to being affected by multiple scattering paths (i.e., non-line-of-sight signals), a dominant path (line-of-sight signal) exists, where the dominant path signal is stronger than the scattering path signal. In this embodiment, the Rician factor is used. The Ricean fading model controls the ratio of the signal power of the main path to that of the scattered path; a larger K-factor indicates that the main path signal dominates. The specific expression is as follows:

[0099] (5);

[0100] in, The Rice factor represents the ratio of the line-of-sight path power to the scattered path power. Indicates a line-of-sight signal. It is path gain. Indicates the scattering path signal. It is Each scattering path gain, It is Delay under each scattering path for Input signal at time, For the index of the scattering path, It is the imaginary unit.

[0101] Combining the Doppler effect and multipath effect described above, we can obtain:

[0102] (6);

[0103] (7);

[0104] Substituting equations (6) and (7) into the Rice fading model of equation (5), then Time-enhanced signal data It can be represented as:

[0105] (8).

[0106] The specific preprocessing steps are as follows: First, combine the I and Q signals from the dataset into a complex signal. Based on the above The expression yields the enhanced signal data. The Rice factor of the channel is updated every certain number of sampling points. and Doppler frequency shift parameters This allows for dynamic changes in the channel, and the variation patterns of these two parameters can be set independently, either linearly or non-linearly. By generating multiple enhanced signals from the original signal in this way, the training framework for the modulation pattern recognition network architecture can obtain richer training data, improving its recognition efficiency and robustness.

[0107] After obtaining enhanced signal data Next, to address the impact of noise on the signal, a Butterworth model is used to design a low-pass filter that effectively removes noise while preserving as much of the original data information as possible. The frequency response of the Butterworth filter can be expressed as:

[0108] (9);

[0109] in, Let be the transfer function of the filter. Represents the independent variable in the frequency domain. This indicates the cutoff frequency of the low-pass filter. Indicates the order of the filter.

[0110] In this embodiment, a 4th-order Butterworth filter is used, and the normalized cutoff frequency is set to 0.2. The resulting filter frequency response is as follows: Figure 2 As shown.

[0111] Enhanced signal The process through the Butterworth filter can be represented as:

[0112] (10);

[0113] in, express Inverse Fourier transform, This represents the signal obtained from vector convolution. This refers to the signal after data preprocessing.

[0114] Figure 3 This is an example of denoising a signal using a low-pass filter. The original signal modulation method is QPSK, and the signal-to-noise ratio is -2. For ease of observation, only 200 points are plotted here. The left side shows the constellation diagram (a), amplitude and phase diagram (b), and spectrum diagram (c) of the original signal. The right side shows 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 constellation diagram of the denoised signal has a more distinct outline, the amplitude and phase changes are more stable, and the spectrum diagram shows better suppression of high-frequency noise.

[0115] For this dataset, k classes are randomly sampled each time to form a subset, i.e., sampling is performed from the task distribution. Then, for each subset, a modulation pattern recognition network architecture is randomly sampled in the search space, and trained using uniform training parameters to obtain the true accuracy of the modulation pattern recognition network architecture on the subset. That is, the task triple (subset, modulation pattern recognition network architecture, true accuracy) is obtained as the training data.

[0116] The mean squared error loss is constructed based on a subset of the cross-scene dataset and the modulation pattern recognition network architecture to train and generate a guided predictor and a performance predictor, thereby obtaining a modulation pattern recognition network architecture applicable to the cross-scene dataset.

[0117] (a5) A mean squared error loss model is constructed based on a subset of the cross-scene dataset and the modulation pattern recognition network architecture to train a guided predictor. The guided predictor is used to guide the scoring network in the latent space to generate the modulation pattern recognition network architecture through diffusion.

[0118] The role of the generative guided predictor is to predict the performance of intermediate results during the denoising process of the diffusion model, guiding the generation of the modulation pattern recognition network architecture in the latent space. After obtaining the cross-scene dataset, the generative guided predictor needs to be trained. During the training phase, the guided predictor is distributed across tasks... This generalization allows the predictor to accurately predict the performance of unknown datasets without additional training. To this end, the predictor includes a dataset encoder to encode and learn relevant features of the dataset.

[0119] The cross-scene dataset consists of triplets of (sub-dataset, modulation pattern recognition network architecture, and true accuracy). During the training of the modulation pattern recognition network architecture search framework, the input to the guided predictor comprises three parts: the sub-dataset, time t, and the latent vector representation of the network architecture after the architecture encoder, followed by a noisy forward diffusion process at time t (i.e., the input to the scoring network). (The noise addition process is referenced.) Figure 4 The forward diffusion process. The output is a prediction with true accuracy.

[0120] Generate a guiding predictor Input-based modulation pattern recognition network architecture and cross-scene datasets A mean squared error loss is constructed to minimize the model's prediction accuracy and true accuracy *s* on each task sampled from the cross-scene dataset distribution. The mean squared error loss is as follows:

[0121] (11);

[0122] in, Indicates time, The modulation pattern recognition network architecture is represented by the architecture encoder, and then after time... The result after forward diffusion with added noise, This is a subset of the sampled source dataset. For prediction accuracy, Modulation pattern recognition network architecture exist The true accuracy on For parameters Regarding the task triplet in the following situations loss function, For the task triple , For the distribution of tasks, This represents the trainable parameters for generating the guided predictor.

[0123] During the deployment and execution of the modulation pattern recognition network architecture search framework, the architecture inputs for generating the guiding predictor are the intermediate results of diffusion model denoising, the denoising time steps, and the features of the target dataset.

[0124] In order to make the generation guide predictor aware of time steps The generator needs to be fed a relevant temporal embedding. First, calculate half of the embedding dimension, then calculate the base of the embedding, then calculate the exponential form of the embedding, and finally, compare the result with the time step. The embeddings are obtained through multiplication, and finally, the sine and cosine embeddings are merged. If the embedding dimension is odd, zero-padding is also required. The generative guide predictor itself consists of convolutional layers, fully connected layers, and activation functions. The temporal embeddings are added after the predictor processes the neural network representation, and then processed by multiple layers. The purpose of the generative guide predictor is to add conditional gradients to each denoising step in the denoising generation process of the diffusion model, so that the denoising process of the scoring network proceeds in the direction that satisfies the desired conditions.

[0125] To enable the cross-domain predictor to be 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 in extracting information relevant to the generator and predictor objectives. This embodiment does not directly input data from the cross-scene dataset, but rather inputs feature embedding representations from the cross-scene dataset. For example, these representations can be extracted from a pre-trained ResNet18 using a feature layer prior to classification prediction. To compress the entire set of instances in the cross-dataset source dataset D into a single latent code z, the ensemble encoder should handle input sets of any size and consistently and agnostively summarize the information to the order of instances, i.e., possess permutation invariance.

[0126] Therefore, the ensemble encoder that generates the guided predictor stacks two permutation-invariant modules with attention-based learnable parameters. The low-level intra-class encoder captures class prototypes that reflect label information, while the high-level inter-class encoder considers the relationships between class prototypes and aggregates them into a latent vector. The ensemble encoder structure used in this embodiment simulates high-order interactions between ensemble elements, enabling the generator and predictor to efficiently extract useful information to achieve each objective.

[0127] Therefore, the design of the dataset encoder in this embodiment is as follows: for a given cross-dataset source dataset... ,in , It is a collection of instances from C categories across the source dataset. It is a collection of labels corresponding to instances in the source dataset across multiple datasets. First, a random sample is taken from a certain category c. ,in It is 3D eigenvectors and For a category containing A collection of data samples. The sampled instances are input into the in-class encoder IntraSetPool, for each class. Encode to obtain class prototype Then, the set representation of each class is fed into the inter-class encoder InterSetPool to generate the dataset representation. As shown below:

[0128] (12);

[0129] (13);

[0130] For the specific architecture of the dataset encoder, an attention-based architecture is mainly adopted. Figure 10 The architecture of the multihead attention module and its derived self-attention block are demonstrated. First, the multihead attention block (MAB) is defined:

[0131] (14);

[0132] (15);

[0133] in, For layer regularization, As a multi-head attention mechanism, For multi-head attention blocks, This is a row-wise feedforward layer, meaning that each instance of the input undergoes independent and identical processing. It is the regularized result of the sum of the input features and the features after multi-head self-attention transformation. Then, it is further processed... A line-by-line feedforward layer was implemented, meaning that each sample underwent independent and identical processing, and then... Add them together.

[0134] Based on a multi-head attention block architecture, the dataset encoder architecture employs existing set-attention blocks (SABs) and multi-head attention pooling (PMA). The SAB uses self-attention to learn the features of each element in the set, while PMA pools the input features into k representative vectors. A set-attention block is an attention-based block, such as... Figure 11 As shown, it makes the characteristics of all instances in the set reflect the relationships between themselves and other instances:

[0135] (16);

[0136] (17);

[0137] in, To collect attention blocks, It is a multilayer perceptron. The multi-head attention mechanism MH(Q,K,V) is used, where the query Q, key K, and value V are all set as the input set X.

[0138] Features encoded from ensemble attention blocks can be learned from PMA in a seed vector. Merge above, by modification Calculate to generate k vectors:

[0139] (18);

[0140] (19);

[0141] in, For multi-head attention pooling.

[0142] Here, k can be of any size; the default setting is k=1 to generate a single latent vector. To extract consistent information independent of the order and size of the input elements, the encoding function should be constructed by stacking permutation equivariant layers E. For each element in set Z, this satisfies any permutation in set Z. The following conditions:

[0143] (20);

[0144] Since all components in both attention-focused blocks and multi-head attention pooling are row computation functions, they are permutationally equivalent by the definition of the equation.

[0145] (a6) A mean squared error loss is constructed to train a performance predictor based on a subset of data points from a cross-scene dataset and a modulation pattern recognition network architecture. The performance predictor is used to predict the true performance of the architecture generated by the estimated diffusion model.

[0146] The performance predictor's role is to predict the performance of the modulation pattern recognition network architecture generated by the diffusion model. It uses the same dataset encoder as the generative guide predictor to encode and learn relevant features from the dataset.

[0147] During the training of the modulation pattern recognition network architecture search framework, the input of the performance predictor consists of two parts: a subset of the dataset and the modulation pattern recognition network architecture. The output is a prediction of the true accuracy.

[0148] Performance predictor Input-based neural network architecture and cross-scene datasets A mean squared error loss is constructed to minimize the model's prediction accuracy and true accuracy *s* on each task sampled from the cross-scene dataset distribution. The mean squared error loss is as follows:

[0149] (twenty one);

[0150] in, This represents the modulation pattern recognition network architecture. This is a subset of the sampled source dataset. For prediction accuracy, Modulation pattern recognition network architecture exist The true accuracy on For parameters Regarding the task triplet in the following situations loss function, For the task triple , For the distribution of tasks, This represents the trainable parameters of the performance predictor.

[0151] The performance predictor consists of a bidirectional graph encoder and a dataset encoder, and the prediction results are output by fully connected layers and activation layers. The bidirectional graph encoder is composed of a forward graph encoder and a backward graph encoder. Both the forward and backward graph encoders are directed acyclic graph encoders, and the difference between them lies in their execution order.

[0152] In a forward graph encoder, for a given neural network architecture ,according to The topological order performs message passing from the predecessor node and iteratively updates the hidden state. :

[0153] (twenty two);

[0154] (twenty three);

[0155] The UPDATE function is a gated loop unit (GRU). It is the first one in the figure. 1 node It is a time step Incoming The node information, the function AGGREGATE is a gate function consisting of mappings and fully connected layers. It is connected to the node The set of preceding nodes, For nodes At any moment The hidden state is represented by, for One-hot encoding of operation types, For nodes At any moment The hidden state is represented by, For set The leading node in.

[0156] For the starting node with an empty leading set Output the zero vector as The hidden states are used, and the last hidden state of the ending node is used as the output of the forward graph encoder. Furthermore, bidirectional encoding can be used 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 is... and the reverse graph encoder The results are connected together and used as the output of the bidirectional graph encoding graph. After being combined with the encoding results of the dataset encoder, the graph passes through a fully connected layer to obtain the performance prediction of the neural network.

[0157] Step 2: Guide the search for a modulation pattern recognition network architecture across datasets, and generate the optimal modulation pattern recognition network architecture based on the given modulation pattern dataset.

[0158] The overall process for generating architectures on the target dataset is as follows: A batch of modulation pattern recognition (MMR) network architectures is generated through a conditional diffusion process, and the performance predictor ranks these architectures. The top k candidate architectures are selected, and their performance is evaluated using existing methods. These architectures and their corresponding performance metrics are added to the fine-tuning training set. The data is weighted according to the performance of the MMR network architectures, and this data is used to fine-tune the generation of the bootstrap predictor and the performance predictor. This process is repeated until the number of architectures with validated performance reaches the expected level. Finally, the architecture with the best performance is selected as the neural network architecture for subsequent training and deployment.

[0159] When the downstream dataset is a target dataset composed of electromagnetic signals transmitted through a channel, the modulation pattern recognition network architecture search framework deployed on the target dataset first defines a conditional diffusion model. After the basic diffusion model is trained, the model parameters are fixed, and then guidance is added to the diffusion process. Inspired by parameterized model guidance schemes, a parameterized generation guidance predictor is added to the generation framework to guide the basic diffusion model to generate an architecture that satisfies a specific objective. Let... We want the neural network architecture to satisfy desired properties (e.g., high accuracy, low latency, high robustness). Then, we solve the following conditional inverse-time SDE to extract the target properties from the features of the target dataset. The information is contained in the distribution function, from the conditional distribution. Generative neural network architecture:

[0160] (twenty four);

[0161] According to Bayes' theorem This allows the conditional fractional function to be decomposed into the sum of two gradients:

[0162] (25);

[0163] The unconditional diffusion model approximates the result through a scoring network. An estimate is needed here. ,because Representing neural network architecture Satisfy target attributes The log-likelihood can be obtained by generating the guided predictor. Fitting:

[0164] (26);

[0165] The introduction of a generative guided predictor is equivalent to guiding the diffusion of the scoring network, identifying whether the network architecture possesses the target attribute by predicting a given modulation pattern. Then, by adding the corresponding gradient to the denoising generation process, modulation pattern recognition network architectures with different attributes can be generated:

[0166] (27);

[0167] in, This is a constant used to determine the strength of the guidance given by the generated guided predictor. Unlike training a regular predictor, the generated guided predictor predicts time steps. Timing Modulation Pattern Recognition Network Architecture The properties of this property are equivalent to the performance of the intermediate result between the modulation pattern recognition network architecture and noise. By combining it with different generative guided predictors, it can be easily adapted to various types of NAS tasks without retraining the basic generative model, making it suitable for various NAS scenarios.

[0168] like Figure 4 As shown, the conditional diffusion process enables rapid deployment of the task on the target dataset. For each category of data in the source dataset, n samples are randomly sampled from each category (e.g., n=20) to serve as inputs for the pre-trained feature extractor and dataset encoder across the scene dataset, eliminating the need for the entire source dataset. The diffusion model then samples from Gaussian noise to begin the denoising process. During this process, the intermediate neural network architecture and the features from the cross-scene dataset are input to the generator-guided predictor. At each step, the performance of the denoising result is predicted, and the gradient of the result with respect to the input is added to the denoising process. In other words, the generator-guided predictor provides conditional information for the diffusion model's denoising process through a method similar to gradient descent, optimizing the denoising result towards higher performance. The final denoised result is input to the architecture decoder, which decodes to obtain the network's connection and operation matrix representations. At this point, the results in the matrix are non-discrete. Through binarization, values ​​≥0.5 are set to 1, and values ​​<0.5 are set to 0, resulting in a 0 / 1 representation of a discrete matrix. Then, based on the definition of the search space, it is determined whether this discrete matrix meets the requirements of the search space, and neural network architectures that do not belong to the search space are excluded, such as those with multiple operations per node, no operations, or non-compliant connections. A batch of compliant neural network architectures and cross-scenario dataset features are then input into the performance ranking predictor. Based on the ranking results of the performance ranking predictor, the architectures ranked higher are selected.

[0169] The above describes how the generation-guided predictor and performance predictor combine a diffusion model to generate modulation pattern recognition network architectures based on a target dataset. Step one provides a good initial value for training the generation-guided predictor and performance predictor. To further improve the performance of the obtained architectures, fine-tuning can be performed on the target dataset. After each architecture generation, based on the ranking results of the performance predictor, the top k modulation pattern recognition network architectures are selected, and their network performance is evaluated using existing evaluation methods to obtain performance estimates for each architecture. These data can then be used to fine-tune the two predictors. Since the training samples are limited, to improve the two predictors' awareness of the architecture representation space, this embodiment updates the two predictors through weighted retraining, thus obtaining a sample-efficient search algorithm. The intuition behind this method is to focus more on high-performance latent networks. During training, the training data is weighted according to the network's performance, and both the generation-guided predictor and performance predictor are fine-tuned on the weighted data distribution. For a fine-tuning training set with known performance... For each architecture Assign a weight This indicates the relative probability that a sample will be sampled during training, and the corresponding loss function is weighted during the training phase. ,in It is the architecture The loss function is used as input. For the weights, rank-based weights are used:

[0170] (28);

[0171] (29);

[0172] in It is a modulation pattern recognition network architecture. It is the architecture The relative weights assigned, It is the number of architectures in the current fine-tuning training set. It is a constant term. It is the architecture The evaluation method It is fine-tuning the training set. express Medium performance is higher than architecture The number of architectures, i.e., architectures The ranking 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.

[0173] Step 3: Train the parameters of the optimal modulation pattern recognition network architecture based on the given modulation recognition dataset to obtain the modulation pattern recognition network model.

[0174] Taking the DeepSig RadioML 2016.10A dataset as an example, this dataset is for wireless communication modulation identification, containing up to 11 modulation types (such as BPSK, QPSK, 16-QAM, etc.) and covering different noise conditions and channel environments. Each sample is a complex sequence of length 128, labeled with the modulation type. The dataset is suitable for tasks such as signal classification, modulation identification, and wireless spectrum sensing using deep learning and traditional machine learning algorithms.

[0175] The dataset is randomly divided into training, validation, and test sets in a 6:2:2 ratio. The data preprocessing procedure follows the same method as in step two for dataset preprocessing for wireless communication modulation identification: the training set is processed using a dynamic simulated channel and low-pass filtering, while the validation set is processed only using low-pass filtering.

[0176] For the modulation pattern recognition network architecture searched during 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.

[0177] The entire training set is denoted as: (30);

[0178] in, For sample subscripts, The total number of samples; For the first A sample, by The sampling points are composed of two IQ channels; For the first The category labels of the samples are: Each category, using Indicates category label The corresponding one-hot vector is used for subsequent training.

[0179] Let the optimal modulation pattern recognition network architecture obtained in the above steps be denoted as . Input Sample Optimal modulation pattern recognition network architecture The resulting prediction is denoted as:

[0180] (31);

[0181] in, yes The probability distribution of each category. The training loss function is the cross-entropy loss, expressed as:

[0182] (32);

[0183] in, Representing the optimal neural network architecture The parameters are then optimized using gradient descent. , represented as:

[0184] (33);

[0185] in, , Respectively and the Model parameters of the optimal modulation pattern recognition network architecture after the next iteration. This represents the learning rate used to train the optimal modulation pattern recognition network architecture. The above process is iterated repeatedly until the performance of the optimal modulation pattern recognition network architecture converges on the validation set, resulting in a modulation pattern recognition model with good performance.

[0186] Step 4: Deploy the modulation pattern recognition model to the channel transmission to perform modulation pattern recognition on the electromagnetic signals transmitted through the channel.

[0187] For a new modulation pattern dataset , Samples in This dataset consists of two streams of I and Q data with N sampling points. This embodiment requires processing this dataset. Each sample in The modulation pattern is identified.

[0188] In this embodiment, the test set generated in step three is used as an example, serving as the input to the optimal modulation pattern recognition model. After the input signal is denoised by the low-pass filter described in step two, it is input into the modulation pattern recognition model obtained in step three, and the resulting prediction is the modulation pattern recognition result of the signal.

[0189] In this embodiment, when the modulation pattern recognition network architecture search framework is trained and deployed to an unknown target dataset, the modulation pattern recognition network architecture search framework does not need to be retrained. Instead, it can directly use the features of the target dataset to generate the network architecture, which greatly reduces the search overhead and reduces the dependence on the amount of data in the features of the target dataset.

[0190] The key aspects of this embodiment lie in the latent space-based scoring network generation and the cross-domain predictor (including a generation-guided predictor and a performance predictor) across scene datasets. The scoring network can be combined with predictors of various task types (accuracy, time delay) without retraining. Furthermore, the modulation pattern recognition network architecture is encoded into the latent space through an architecture encoder-decoder, reducing the difficulty of learning the distribution of the modulation pattern recognition network architecture. The combination of the scoring network and the cross-domain predictor models the distribution of the search space. The cross-scene dataset predictor only requires one training iteration to be deployed on different datasets. The introduction of the dataset encoder plays a crucial role. The cross-domain predictor learns the performance of the architecture on specific tasks through dataset features, thereby achieving prediction and generation across scene datasets.

[0191] The advantage of this embodiment is that other search methods require retraining when deploying searches on various downstream datasets, while this embodiment only needs to pre-train the architecture encoder-decoder, diffusion model, and cross-domain predictor across scene datasets once before direct deployment to various downstream datasets. This avoids additional search overhead, significantly accelerating search efficiency and generating high-performance modulation pattern recognition network architectures. Furthermore, this embodiment does not require training on the entire downstream dataset during deployment, reducing data dependency and making it suitable for scenarios with limited data. This embodiment is also scalable; by simply adjusting the prediction objective of the performance predictor based on the cross-scene dataset, modulation pattern recognition network architectures with different properties can be generated, including high accuracy, low latency, high robustness, hardware metrics, etc., enabling the search of networks for specific tasks and achieving dedicated networks for specific needs.

[0192] This invention provides a cross-dataset generation framework in which the specific network architecture can be modified, including but not limited to the selection of the search space, the architecture design of the architecture encoder and decoder, the architecture design of the diffusion model, the architecture design of the generation guide predictor and the performance predictor, and the usage method and architecture design of the dataset encoder. For specific tasks and datasets, suitable deep learning architectures can be selected for replacement, which has good scalability.

[0193] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A modulation pattern recognition method based on generative neural network architecture search, characterized in that, include: A modulation pattern recognition network architecture search framework is trained to obtain a generative model, a generative guidance predictor, and a performance predictor; Guide the search for modulation pattern recognition network architecture across datasets, and generate the optimal modulation pattern recognition network architecture based on the given modulation pattern dataset. The optimal modulation pattern recognition network architecture is trained with parameters based on the given modulation recognition dataset 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 signals transmitted through the channel. The training process of the modulation pattern recognition network architecture search framework is as follows: A parameterized search space for modulation pattern recognition network architecture is constructed based on the modulation pattern recognition task. Each modulation pattern recognition network architecture in the parameterized search space is regarded as a directed acyclic graph. The nodes of the graph represent operations, and the directed edges represent the direction of information transmission. Each pattern recognition network architecture is represented as an adjacency matrix and an operator operation type matrix. Train the network architecture encoder-decoder in the parameterized search space, and encode the modulation pattern recognition network architecture into the latent space using the network architecture encoder-decoder; The diffusion model is trained in the latent space, and the back-diffusion process of the diffusion model is used as the generative model of the modulation pattern recognition network architecture, the performance predictor of the modulation pattern recognition network architecture, and the generative guidance predictor of the modulation pattern recognition network architecture. 5-1) Training a generative guided predictor for a modulation pattern recognition network architecture in the latent space, specifically: Construct a cross-scenario dataset, including a subset of datasets, a modulation pattern recognition network architecture, and true accuracy triples; The modulation pattern recognition network architecture representation matrix is ​​input into the architecture encoder for encoding to obtain the latent space representation vector of the architecture; Based on the forward diffusion process, select the time. , to diffuse and add noise to the latent space vector; The subset of data and the noisy latent space vectors are input to generate a guided predictor, which predicts the corresponding true accuracy. Based on the prediction accuracy and the true accuracy, the mean squared error loss is constructed to train and generate the guided predictor model. During the deployment and execution of the modulation pattern recognition model, the architectural inputs for generating the guiding predictor are the intermediate results of the diffusion model denoising, the denoising time steps, and the features of the target dataset. 5-2) Training the performance predictor of the modulation pattern recognition network architecture, specifically: Construct a cross-scenario dataset, including a subset of datasets, a modulation pattern recognition network architecture, and true accuracy triples; Input the performance predictor with the subset of data and the modulation pattern recognition network architecture representation matrix to predict the corresponding true accuracy. Based on the predicted accuracy and the true accuracy, construct the mean squared error loss to train the performance predictor model. During the deployment and execution of the modulation pattern recognition model, the input to the performance predictor is the modulation pattern recognition network architecture finally generated by the diffusion model denoising and the features of the target dataset.

2. The modulation pattern recognition method based on generative neural network architecture search according to claim 1, characterized in that, The network architecture encoder and decoder are trained in the parameterized search space of the modulation pattern recognition network architecture, as follows: Noise is added to the modulation pattern recognition network architecture representation matrix to train the architecture encoder to encode the modulation pattern recognition network architecture into a latent representation vector. Noise is added to the latent representation vector and then input into the architecture decoder. A reconstruction loss is constructed based on the output modulation pattern recognition network architecture and the input modulation pattern recognition network architecture to train the architecture encoder and decoder.

3. The modulation pattern recognition method based on generative neural network architecture search according to claim 1, 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 pattern recognition network architecture. The modulation pattern recognition network architecture in the parameterized search space is gradually noise-added through the forward diffusion process, and the distribution of the modulation pattern recognition network architecture is mapped to the known prior distribution to obtain the noisy architecture hidden representation. We define an inverse stochastic differential equation to describe the unconditional diffusion process of the modulation pattern recognition network architecture, and denoise the hidden representation of the noisy architecture through the reverse diffusion process. In backdiffusion, the trainable parameters of the scoring network are trained to approximate the score function of the unknown terms in the backdiffusion process, thereby obtaining the trained generative model.

4. The modulation pattern recognition method based on generative neural network architecture search according to claim 3, characterized in that, The optimized representation of the trainable parameters in the generative model is as follows: in, These are the trainable parameters after training the scoring network. These are the trainable parameters of the scoring network. Indicates the moment of the diffusion process. Indicates taking the expected value. Indicates time-based Given a positive weighting function, At time 0, i.e., the function mapping of the initial modulation pattern recognition network architecture, For a moment Function mapping of time-modulation pattern recognition network architecture Indicates time 0 expectations, Indicates that in a given In the case of time hour expectations, Indicates and time For the input scoring network, For a fractional function with unknown terms, For about gradient, For time-based of The marginal probability distribution.

5. The modulation pattern recognition method based on generative neural network architecture search according to claim 1, characterized in that, The cross-scenario dataset is constructed as follows: The dataset for wireless communication modulation recognition is preprocessed to obtain the source dataset. Each time, k categories are randomly sampled from the source dataset to form a subset. A modulation pattern recognition network architecture is randomly sampled in the parameterized search space. The true accuracy of the modulation pattern recognition network architecture on the subset is trained. The task triple consisting of the subset, the modulation pattern recognition network architecture, and the true accuracy is used as cross-training data to construct a cross-scene dataset. The mean squared error loss is constructed based on a subset of the cross-scene dataset and the modulation pattern recognition network architecture to train a cross-domain predictor to search for the modulation pattern recognition network architecture of the source dataset. The generation model generates the optimal modulation pattern recognition network architecture based on the task dataset. The cross-domain predictor includes a generation guidance predictor and a performance predictor.

6. The modulation pattern recognition method based on generative neural network architecture search according to claim 5, characterized in that, The process of constructing the source dataset is as follows: Dynamic Simulated Channel: The Doppler and multipath effects are introduced to simulate the impact of real channels on signals, thus constructing a dynamic simulated channel. Data augmentation: The I and Q signals in the dataset are combined into an input complex signal. After the input complex signal passes through a dynamic simulation channel, the enhanced signal data is obtained. By updating the K factor and Doppler frequency shift parameters of the channel at several sampling points, the dynamic change of the channel is realized. Filtering: The enhanced signal data is filtered using a low-pass filter, and the processed data is used as the source dataset. The expression for the dynamic simulated channel is as follows: in, for The signal data after real-time enhancement Rice factor, For Doppler frequency shift parameters, for Input signal at time, for Input signal at time, For path gain, The number of scattering paths, It is the first Each scattering path gain, It is the first Delay under each scattering path It is the imaginary unit.

7. The modulation pattern recognition method based on generative neural network architecture search according to claim 1, characterized in that, Both the generation-guided predictor and the performance predictor are equipped with a dataset encoder, and the two dataset encoders have the same structure. For the generative guided predictor, the mean squared error loss constructed from the prediction accuracy and the true accuracy is as follows: in, Indicates time, The modulation pattern recognition network architecture is represented by the architecture encoder, and then after time... The result after forward diffusion with added noise, This is a subset of the sampled source dataset. For prediction accuracy, Modulation pattern recognition network architecture exist The true accuracy on For parameters Regarding the task triplet in the following situations loss function, For the task triple , For the distribution of tasks, This represents the trainable parameters for generating the guided predictor; For the performance predictor, the mean squared error loss constructed from the prediction accuracy and the true accuracy is as follows: in, For prediction accuracy, For parameters Regarding the task triplet in the following situations loss function, Indicates performance predictor Trainable parameters.

8. The modulation pattern recognition method based on generative neural network architecture search according to claim 1, characterized in that, A guided predictor and a performance predictor are trained using a cross-scene dataset. The guided predictor is used to guide the scoring network to generate the latent space representation of the modulation pattern recognition network architecture through the diffusion model in the latent space. The architecture decoder maps the latent space representation generated by the diffusion model to the modulation pattern recognition network architecture in the original search space. The performance predictor is used to predict the performance of the mapped modulation pattern recognition network architecture. During the deployment and execution of the modulation pattern recognition network architecture search framework, the generative model is guided by the generation guide predictor to generate a set of modulation pattern recognition network architectures applicable to a given modulation recognition dataset. The set of modulation pattern recognition network architectures is sorted in descending order based on the performance predictor. The modulation pattern recognition network architectures with the highest performance ranking are selected for performance evaluation. The generation guide predictor and the performance predictor are fine-tuned using the selected modulation pattern recognition network architectures and their corresponding evaluation accuracy to adapt to the generation of the given modulation recognition dataset.

Images