Cross-domain radiation source individual identification method based on multi-view neural architecture search

Abstract

The invention discloses a cross-domain radiation source individual identification method based on multi-view neural architecture search, which combines a neural architecture search and multi-view representation fusion solution, designs a neural architecture search method under a signal multi-view, sets a plurality of branches to process signal features under different views, and improves the identification accuracy of the cross-domain radiation source individual. According to the method, the multi-view features are extracted, an architecture building block with the most cross-domain individual recognition capability for each signal feature is searched, efficient integration of the multi-view features is achieved through dynamic updating of probability weights, the performance and generalization capability of the model in different scenes are remarkably improved, and the method is suitable for radiation source individual recognition tasks in the multiple scenes.

Description

Technical Field

[0001] The present invention belongs to the technical field of radiation source individual identification, and mainly relates to a cross-domain radiation source individual identification method based on multi-view neural architecture search. Background Art

[0002] Radiation source individual identification, also known as Specific Emitter Identification (SEI), is a technology that uses radio frequency fingerprints generated by hardware non-ideal characteristics to distinguish different radiation source targets. As the modulation contained in radiation source signals becomes more and more complex, the robustness of radiation source individual identification under different modulation environments decreases significantly. With the increasingly complex electromagnetic environment and the overlap between pulse signals, using traditional single conventional parameter features to identify radiation sources individually can no longer meet the requirements of accuracy, time, and storage space. How to improve the completeness of features and combine the advantages of different signal features has become one of the challenges in the task of radiation source individual identification.

[0003] As an important model in machine learning, neural networks are increasingly applied to the field of radiation source individual identification due to their excellent feature extraction and representation capabilities. For neural networks, reasonable architecture design can effectively capture important features in data, thereby improving the performance of the model. However, designing an efficient neural network structure often requires profound professional knowledge and a large amount of experimental work, which makes network architecture design a complex and time-consuming task.

[0004] The Naval Aviation University of the Chinese People's Liberation Army disclosed an unsupervised domain adaptation radiation source individual identification method based on the independence criterion in Patent CN119046797A. Although this method solves the problem of domain adaptation in the field of radiation source individual identification to a certain extent, using a fixed network architecture as a feature extractor often results in performance fluctuations in the face of complex and changing electromagnetic environments; the Naval Aviation University of the Chinese People's Liberation Army disclosed a radar radiation source individual identification method and device based on multi-dimensional features in Patent CN118861624A, but the feature-level fusion used in this patent may lose some useful information, affecting subsequent identification accuracy. The University of Electronic Science and Technology disclosed an intelligent radar radiation source individual identification method based on multi-branch feature fusion in Patent CN118465721A, which uses multi-branch processing of multiple features and combines an AFF feature fusion module for feature fusion. However, constructing a feature extraction network requires a large amount of labor cost and time consumption. Summary of the Invention

[0005] In view of the problem that the performance of individual identification of radiation sources degrades in cross-domain scenarios in the prior art, the present invention provides a cross-domain radiation source individual identification method based on multi-view neural architecture search. By combining a solution of neural architecture search and multi-view feature fusion, a neural architecture search method under multiple views of signals is designed. Multiple branches are set to process signal features under different views, and architecture building blocks with the most individual cross-domain identification ability for each signal feature are searched. Moreover, the dynamic update of probability weights is adopted to achieve the efficient integration of multi-view features, significantly improving the performance and generalization ability of the model in different scenarios and being applicable to the radiation source individual identification task in multiple scenarios.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows: A cross-domain radiation source individual identification method based on multi-view neural architecture search, comprising the following steps:

[0007] S1. Multi-view data preparation: Collect radio signals, perform pulse detection through the sliding window method after preprocessing, and divide samples; use different feature extraction methods for radio signal samples to obtain different features; the feature extraction methods at least include discrete wavelet transform to obtain the decomposition and reconstruction features of the signal, and short-time Fourier transform to analyze the time-frequency domain features of the signal.

[0008] S2. Design the search space based on the Inception structure: The search space at least includes Inception modules. Each Inception module includes a normal block and a reduction block. The normal block contains 4 branches, and the number of output channels is the same as the number of input channels; the reduction block contains 2 branches, and the number of output channels is halved. The search space contains a search operation pool, and the selectable operations are not unique; an optional operation is placed on each branch in the Inception module, and filters with different kernel sizes are used and their outputs are concatenated along the channel dimension to capture local and global features simultaneously.

[0009] S3. Model search: Based on the neural architecture search method optimized by gradient, set multiple branches to input different features respectively, search the feature extraction networks of each feature respectively, optimize the search network through gradient descent, and obtain the optimal architecture building blocks; set the probability fusion weights, and dynamically update the fusion weights during the search process to achieve the integration of multi-view information.

[0010] S4. Domain adaptation: According to the optimal architecture building blocks obtained by searching in step S3, stack the searched normal blocks and reduction blocks 8 times to build a complete network. The reduction blocks are placed on odd layers, and the normal blocks are placed on even layers. The output of each previous module is used as the input of the next module. Perform domain adaptation on the built complete network through the labeled data in the source domain and the unlabeled data in the target domain, and verify the individual identification performance of the model in the target domain.

[0011] As an improvement of the present invention, the discrete wavelet transform in step S1 is implemented based on the continuous wavelet transform, and the continuous wavelet transform is defined as:

[0012]

[0013] WT x (a, b) represents the projection of the signal x(t) on the wavelet basis function ψ(t). The discrete wavelet transform discretizes the scale factor a and the translation factor b to obtain a set of discrete wavelet functions:

[0014]

[0015] Discrete wavelet transform coefficients The calculation is as follows:

[0016]

[0017] where a0 is the scale base value for controlling the stretching step of the wavelet, j is the scale parameter for controlling the width and frequency of the wavelet, b0 is the translation base value for controlling the translation step of the wavelet, k is the translation parameter, and kb0 represents the translation amount of the wavelet on the time axis. represents the wavelet function after scale stretching and time translation. is the normalization factor to ensure the wavelet energy is consistent at different scales.

[0018] As an improvement of the present invention, the short-time Fourier transform in step S1 to analyze the time-frequency domain characteristics of the signal specifically includes the following steps:

[0019] S11: Select the Hamming window as the window function, and the window function is defined as:

[0020]

[0021] where T w is the total window length. If the signal sampling rate is f s , and the number of signal samples is N, then

[0022] S12: Divide the original signal into overlapping or non-overlapping short time segments and weight them through the window function:

[0023] x w (t) = x(t) · w(t)

[0024] S13: Apply the Fourier transform to the weighted signal in each time segment, convert the signal from the time domain to the frequency domain, and calculate the spectral information of each time segment:

[0025]

[0026] where X(ω,t) is the spectral component of the signal at time t and frequency ω, and x w (t) is the signal obtained by windowing the original signal x(t) with the window function w(t), and x w (t) = x(t)·w(t - t0).

[0027] As another improvement of the present invention, the mathematical representation of the Inception module in step S2 is specifically as follows:

[0028] y = concat([conv(x, W f ) for f in F])

[0029] z = conv(y, W b )

[0030] y out = Relu(batchnorm(z))

[0031] where conv(·) represents the convolution operation, W f and W b represent the convolution kernels, batchnorm(·) represents batch normalization, Relu(·) is the activation function, F represents the search operation pool, traverse each operation f therein, perform convolution operations on them, then splice the obtained feature representations by concat(·) by channels, then calculate the 1×1 convolution, and then obtain the module output through batch normalization and the activation function.

[0032] As another improvement of the present invention, in step S2, the defined search operation pool includes ten different operations: 3×1 convolution, 5×1 convolution, 7×1 convolution, 9×1 convolution, 11×1 convolution, 3×1 max pooling, 5×1 max pooling, 3×1 average pooling, 5×1 average pooling, and skip connection. Each branch in the Inception module places an optional operation that is one of the ten operations.

[0033] As yet another improvement of the present invention, in the neural architecture search method based on gradient optimization in step S3, it includes double - layer optimization. The double - layer optimization is divided into upper - layer optimization and lower - layer optimization. The architecture parameter α is used as the upper - layer variable, and based on the loss L val on the validation set, the gradient is calculated through backpropagation and optimized using gradient descent; the network weight w is used as the lower - layer variable, and based on the loss L train on the training set, the optimal network weight is obtained by minimizing the training error. The double - layer optimization is specifically as follows:

[0034]

[0035] After the search is completed, according to the maximum weight of the architecture parameters, the optimal discrete operation between each node is selected to generate the final network architecture:

[0036]

[0037] Among them, is the architecture parameter array of all optional operations between node i and node j, and the finally selected operation o is obtained by taking argmax (i,j) .

[0038] As another improvement of the present invention, the class probability classification after dynamically updating the probability weights during the search process in step S3 is specifically:

[0039] P = γ1·P1 + γ2·P2 + … + γ N ·P N

[0040] Among them, γ1, …, γ N are the set dynamic fusion weights, and P i are the respective features extracted from the input data of each branch by their respective feature extractors, and are the corresponding probability distributions obtained after passing through the subsequent fully connected layer and softmax.

[0041] As a further improvement of the present invention, the loss function in the model search in step S3 includes cross-entropy loss and domain adaptation loss based on maximum mean discrepancy. The cross-entropy loss is specifically:

[0042]

[0043] Among them, N is the number of samples, y n,i is the actual label of the nth sample for the ith class, and p n,i is the probability that the nth sample is predicted to be the ith class;

[0044] The domain adaptation loss based on maximum mean discrepancy is specifically:

[0045]

[0046] Among them is the source domain feature, is the target domain feature, m S and m T are the total number of samples in the source domain and the target domain respectively, and φ(·) maps each sample to the Hilbert space related to the kernel ;

[0047] The final combined loss function is:

[0048] L total = L clf + λLMMD

[0049] Among them, λ is a parameter reflecting the influence degree of the domain adaptation loss.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention proposes a cross-domain radiation source individual recognition method based on multi-view neural architecture search. Through a variety of feature extraction algorithms, the feature expressions of signals from different perspectives are obtained; multiple branches are set, and the most suitable network architecture is dynamically searched for each branch, so as to fully explore and utilize the advantages of features from each perspective; a joint loss function is designed, which combines the classification loss with the domain adaptation loss based on the maximum mean discrepancy (MMD), effectively guiding the model optimization and searching for a network architecture with stronger domain adaptation ability; at the same time, a dynamic probability weight fusion mechanism is also adopted to weight and optimize the prediction results of multiple branches, further improving the robustness and generalization ability of the model. The finally generated high-performance network architecture has both good scalability and practicability, providing an efficient and stable solution for the adaptation problem in the field of radiation source signal individual recognition. The method of the present invention combines multi-view feature representation fusion and neural architecture search, effectively combines the advantages of different features, and improves the individual recognition performance and flexibility of the model in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is the step flow chart of the method of the present invention;

[0052] Figure 2 is the schematic diagram of the model structure in the model search stage of the method of the present invention

[0053] Figure 3 is the schematic diagram of the model structure in the domain adaptation stage of the method of the present invention;

[0054] Figure 4 is the schematic diagram of the structure of the Inception module in the method of the present invention, where

[0055] Figure 4 (a) is the schematic diagram of the ordinary block structure;

[0056] Figure 4 (b) is the schematic diagram of the reduction block structure;

[0057] Figure 5 is the schematic diagram of the probability weight fusion structure in Embodiment 2 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] The present invention will be further clarified below in conjunction with the drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0059] Embodiment 1

[0060] A cross-domain radiation source individual recognition method based on multi-view neural architecture search. The flowchart of the steps of the method of the present invention is as shown in Figure 1 and specifically includes the following steps:

[0061] Step S1, multi-view data preparation: Collect radio signals, perform pulse detection by the sliding window method after preprocessing, divide samples, and apply different feature extraction methods to obtain different features; the feature extraction methods at least include discrete wavelet transform to obtain the decomposition and reconstruction features of the signal, and short-time Fourier transform to analyze the time-frequency domain features of the signal. The network sets multiple input branches to process different input features obtained by using different feature extraction methods respectively.

[0062] Step S2, search space construction: The search space includes two types of Inception modules, the normal block and the reduction block. The schematic structural diagrams of the two modules are as shown in Figure 4 where Figure 4 (a) is the schematic structural diagram of the normal block, Figure 4 (b) is the schematic structural diagram of the reduction block. The normal block contains 4 branches, and the number of output channels is the same as the number of input channels; the reduction block contains 2 branches, and the number of output channels is halved. An optional operation is placed on each branch of the module. Construct a search operation pool containing various operations for selecting the operations of the module branches, including convolution operations with various kernel sizes, max pooling, average pooling, skip connections, etc., to represent the diversity of the network architecture.

[0063] The search space designed in the present invention is based on the Inception structure. Inception contains a multi-branch structure, and different branches can select different operations from the search operation pool, providing a broad search space for searching different branch operations. The Inception module can enable the network to capture multi-scale and multi-resolution features, and can effectively improve the performance of the deep neural network. The core idea of the Inception module is to use multiple filters with different kernel sizes and concatenate their outputs, enabling the network to capture local and global information simultaneously. The principle of the Inception module is as follows:

[0064] If x is the input of the module and F is the set of filter kernel sizes to be used in the Inception block. For each filter kernel size f in F, convolve the input feature map x with a set of filters W f to obtain the output feature map y f which is: y f = conv(x, W f ), and then concatenate the outputs of all branches along the channel dimension to obtain the final output feature map: y = concat(y f1 , y f2,...,y fn )。To further reduce the number of channels and control the computational complexity, the concatenated feature map is passed through a bottleneck layer, which performs 1×1 convolution using filters with a smaller kernel size. After this layer, a ReLU activation function and a batch normalization layer are connected. The mathematical representation of the Inception module is as follows, W b is the set of filters for each branch, and the output y is calculated out :

[0065] y = concat([conv(x, W f ) for f in F])

[0066] z = conv(y, W b )

[0067] y out = Relu(batchnorm(z))

[0068] Two types of Inception modules proposed in the present invention are used for architecture construction: the normal block and the reduction block. The input and output dimensions of the normal block are the same, and it contains 4 branches; the output dimension of the reduction block is halved, and it contains 2 branches. The proposed search operation pool includes 3×1 convolution, 5×1 convolution, 7×1 convolution, 9×1 convolution, 11×1 convolution, 3×1 max pooling, 5×1 max pooling, 3×1 average pooling, 5×1 average pooling, and skip connections. Each branch can be one of these ten operations. In both the search stage and the training stage, Inception modules are stacked to form the construction of the network architecture.

[0069] Step S3, model search stage: Design a multi-perspective neural architecture search method, optimize the network architectures of each branch respectively through gradient descent, and learn high-performance architecture building blocks with stronger feature extraction ability for this input feature; propose a joint loss function combining cross-entropy loss and domain adaptation loss to guide the optimization of model architecture parameters and network parameters; set probability fusion weights to fuse the prediction results of each branch, and dynamically update the probability weights during the search process, so as to effectively integrate multi-perspective information. The optimal architecture building blocks and the optimal probability fusion weights will be obtained in the search stage.

[0070] The structural schematic diagram of the model search stage is as Figure 2The search for a single Inception normal block is shown in the figure. The rectangles containing numbers represent different nodes, and the arrows of different colors represent different optional operations. To make the search space continuous, the initial model includes all optional operations and assigns architecture weights to each operation. The size of the weights is reflected by the thickness of the arrows. The architecture parameters are optimized through a two-layer optimization. The softmax function is applied to the architecture weights, and finally, the operation with the highest probability is selected as the final chosen operation. While searching for the architecture parameters, the optimal fusion weight γ is explored. The fusion weight is dynamically assigned according to the importance of the features, and then the features from different perspectives are fused with dynamic probability weights.

[0071] In the model search stage, based on the gradient-based neural architecture search method DARTS, a multi-perspective neural architecture search method is designed. Multiple branches are set to input different features respectively, and the feature extraction networks for each feature are searched separately. The search network is optimized through gradient descent to learn high-performance architecture building blocks that adapt to the differences between the source domain and the target domain. A joint loss function that includes the cross-entropy loss and the domain adaptation loss based on the maximum mean discrepancy is proposed to guide the model optimization. A dynamic probability weight fusion method is proposed to fuse the prediction results of each branch according to the dynamic weights, achieving an effective integration of multi-perspective information.

[0072] DARTS (Differentiable Architecture Search) is a gradient-based neural network architecture search (NAS) method. It significantly improves the search efficiency by transforming the discrete architecture search problem into a continuous optimization problem. Compared with traditional NAS methods (such as reinforcement learning or evolutionary algorithms), DARTS avoids the expensive search cost and enables a more efficient architecture search. The process of searching for architecture building blocks is as follows:

[0073] Each candidate architecture in DARTS consists of several nodes and the operations between the nodes. The nodes represent feature maps, and the operations between the nodes are defined as a series of transformations (such as convolution, pooling, etc.) from the input node to the output node. Suppose O is the set of all optional operations in the search operation pool, and each operation o ∈ O is a function that acts on the input feature map x (i) . DARTS represents the selection of the operation between nodes as a probability-weighted mixed operation, which has the following form:

[0074]

[0075] is a learnable parameter associated with operation o, representing the weight of the operation. By calculating the softmax probability for all operations, the selection of each hybrid operation can be optimized in a continuous probability space. The DARTS algorithm maps architecture search from a discrete space to a continuous space, allowing the search process to be solved directly through gradient optimization without the need for independent training and validation of each candidate architecture.

[0076] After the search is complete, each blend operation will be replaced by the discrete operation with the highest probability. Specifically, for each pair of nodes (i, j), the operation with the largest weight o is selected. (i,j) As the final discrete operation:

[0077]

[0078] The discretization method ensures the final architecture acquisition and search efficiency.

[0079] The optimization of the model architecture is achieved through two-layer optimization, where the architecture parameter α is the upper-layer variable that determines the architecture structure; the network weight w is the lower-layer variable that determines the optimization of the network's learnable parameters during training. The goal of architecture search is to find the model that minimizes the validation loss. α * , where the architecture-dependent weights w * By minimizing the training loss The two-layer optimization is as follows:

[0080]

[0081] Directly solving the above two-level optimization problem is computationally expensive. DARTS uses a gradient-based approximation method. In each optimization, by approximately calculating the gradient of the upper-level variable α, it is possible to avoid directly solving the lower-level optimization problem. The expression of the approximate gradient is:

[0082]

[0083] Where η is the learning rate, which represents the change after a single step of gradient descent on the weight w. This method utilizes the gradient information of the weight w with respect to the training loss, effectively reducing the computational complexity.

[0084] The method presented in this paper combines multi-view representation fusion with DARTS, creating multiple independent branches, each responsible for processing feature inputs from a different viewpoint. Within each branch, the DARTS approach dynamically optimizes the network architecture that best suits the input features of that branch, ensuring that the diversity and specificity of features from different viewpoints are fully explored and utilized. This design effectively leverages the advantages of multiple viewpoint features, enhancing the model's expressive power and domain adaptability.

[0085] The loss function in this step of model search includes cross-entropy loss and domain adaptation loss based on maximum mean discrepancy. The cross-entropy loss is specifically as follows:

[0086]

[0087] where N is the number of samples, y n,i is the actual label of the i-th class of the n-th sample, and p n,i is the probability that the n-th sample is predicted as the i-th class;

[0088] The domain adaptation loss measures the difference between two domain distributions using the maximum mean discrepancy method. It finds a function in the function space that takes large values on one set of samples and small values on another set of samples, thereby maximizing the expected difference between the two sets of samples.

[0089] Given the source domain samples X s ={x1, x2, …, x m} and the target domain samples X t ={x1, x2, …, x n}, let the distribution of the source domain sample set be p s , and the distribution of the target domain sample set be p t . There is usually a distribution shift between the data in the source domain and the target domain, that is, p s ≠p t [[ID=E34]]. Through the maximum mean discrepancy (MMD), this distribution shift is quantified and used as the optimization objective to improve the generalization performance of the model on the target domain. The definition of MMD is:

[0090]

[0091] where H is the reproducing kernel Hilbert space, and sup means finding a function f in the function set F such that has the maximum value.

[0092] Given the source domain features and the target domain features , the domain adaptation loss based on the maximum mean discrepancy is:

[0093]

[0094] where m s and m T are the total numbers of samples in the source domain and the target domain respectively, and φ(·) maps each sample to the Hilbert space related to the kernel . This MMD value is optimized in the direction of becoming smaller as the domain adaptation loss. The final combined loss function is L total =L clf +λLMMD 。

[0095] In the subsequent stage of the model, a dynamic probability weight fusion method is adopted to perform weighted fusion on the prediction results of each branch. The weight allocation is dynamically adjusted according to the contribution degree of each branch to the final prediction result, so as to achieve efficient fusion of features from different perspectives.

[0096] A dynamic probability weight fusion method is proposed to effectively combine the advantages of each feature, improve the recognition accuracy and generalization ability. Multiple branches are set to input different features respectively, and the optimal fusion weights are dynamically searched in the search stage. This weight reflects the contribution degree of the features of each branch to the result, and the dynamic allocation of the weight makes the model more flexible and able to adapt to the characteristics of different tasks. Specifically, it includes: assuming that N multi-branch input data are X1,…,X N 。After passing through the feature extractors searched respectively, the features Z1,…,Z after feature extraction are obtained N :

[0097] Z i =Inceptioni(X i ), i = 1, 2, 3

[0098] Followed by a fully connected layer and softmax to obtain the corresponding probability distribution:

[0099] P i =Softmax(FC(Z i ))), i = 1, 2, 3

[0100] Set the dynamic fusion weights γ i ,…, γ N , and perform dynamic optimization based on a two-layer optimization method. The final class probability distribution P = γ1·P1 + γ2·P2 + … + γ N ·P N 。

[0101] Step S4, domain adaptation stage: Based on the optimal network building blocks determined in the search stage, set the stacking layer number to 8, place the reduction blocks on the odd layers and the normal blocks on the even layers, and use the output of the previous module as the input of the next module to stack and form a complete network architecture; train the constructed complete network with the labeled data in the source domain and the unlabeled data in the target domain.

[0102] The structure diagram of the domain adaptation stage is as Figure 3As shown. After the search stage of step S3, the optimal architecture building blocks of each branch and the probability fusion weights most suitable for the current task are obtained. At this time, only one color of arrow is included between each node of the Inception module, representing the determination of operation selection. The ordinary blocks and reduction blocks obtained by the search are stacked 8 times to build a complete network. The input source domain has labeled data and the target domain has unlabeled data, which are trained from scratch and verified on the target domain test set.

[0103] Embodiment 2

[0104] A cross-domain emitter individual recognition method based on multi-view neural architecture search, including multi-view data preparation, search space design, model search stage and domain adaptation stage, specifically including the following steps:

[0105] Step S1: Multi-view data preparation, including the production of raw data and multi-view feature extraction.

[0106] The raw data used in this embodiment is a radio dataset produced in the laboratory, the signal type is pulsed Doppler, and IQ two-channel acquisition is performed. The data contains 7 individuals (transmitters) and 8 modulation methods, namely linear frequency modulation (LFM), non-linear frequency modulation (NLFM), binary frequency shift keying (BFSK), quadrature frequency shift keying (QFSK), binary phase shift keying (BPSK), quadrature phase shift keying (QPSK), binary frequency shift combined binary phase modulation (BFSK_BPSK), and quadrature frequency shift combined binary phase modulation (QFSK_BPSK). After data acquisition, after preprocessing such as denoising and normalization, pulse detection is performed by the sliding window method, and it is divided into multiple samples according to a length of 200. Using the emitter individual as the label, a dataset is constructed. The dataset is divided into the source domain and the target domain according to the modulation method. The divided domains are FM (including modulations LFM and NLFM), FSK (including modulations BFSK and QFSK), PSK (including modulations BPSK and QPSK), and FSK_PSK (including modulations BFSK_BPSK and QFSK_BPSK).

[0107] The discrete wavelet transform is used to obtain the decomposition and reconstruction features of the signal. The advantage of the discrete wavelet transform method is that it can decompose non-linear and non-stationary signals at multiple scales. The signal is decomposed into a series of scale coefficients (approximate coefficients) and wavelet coefficients (detail coefficients) through the discrete wavelet transform. It can effectively extract useful information from the signal. The discrete wavelet transform is implemented on the basis of the continuous wavelet transform. If the initial signal is x(t), the wavelet basis function or mother wavelet ψ(t) is defined as:

[0108]

[0109] Among them, a represents the scale factor and b represents the translation factor. The dilation of the wavelet basis function is characterized by the scale factor a. Therefore, the scale factor a determines the analysis interval of the wavelet transform.

[0110] The continuous wavelet transform is defined as:

[0111]

[0112] WT x (a, b) represents the projection of the signal x(t) on the wavelet basis function ψ(t). The discrete wavelet transform discretizes the scale factor a and the translation factor b to obtain a set of discrete wavelet functions:

[0113]

[0114] The discrete wavelet transform can be calculated based on the following formula:

[0115]

[0116] The short-time Fourier transform is used to analyze the time-frequency domain characteristics of the signal, which is achieved by dividing the signal in time and performing the Fourier transform on each segment of the signal. Specifically, it includes:

[0117] Select the Hamming window as the window function, and the window function is defined as:

[0118]

[0119] The original signal is divided into overlapping or non-overlapping short time segments and weighted by the window function: x w (t) = x(t) · w(t).

[0120] Apply the Fourier transform to the weighted signal in each time segment to transform the signal from the time domain to the frequency domain.

[0121] Calculate the spectral information of each time segment.

[0122]

[0123] Finally, in this embodiment, the original signal features, discrete wavelet transform features, and short-time Fourier transform features form more representative multi-perspective features as the model input.

[0124] Step S2: Search space design, design the search space based on the Inception structure.

[0125] This search space makes full use of the characteristics of the multi-branch structure, providing rich operation options for each branch, thus constructing a vast search space. The Inception module effectively improves the performance of deep neural networks through multi-scale and multi-resolution feature extraction. Its core idea is to use filters with different kernel sizes and concatenate their outputs along the channel dimension to capture local and global features simultaneously. The principle of the Inception module is as follows:

[0126] Assume the module input is x and the set of filter kernel sizes is F. For each kernel size f ∈ F, the input x is convolved through the filter W f The output feature map is:

[0127] y f = conv(x, W f )

[0128] The outputs of all branches are concatenated along the channel dimension to obtain the concatenated feature map:

[0129] y = concat(y f1 , y f2 ,..., y fn )

[0130] To further reduce the channel dimension and control the computational complexity, a bottleneck layer is used to perform 1×1 convolution on the concatenated feature map. After the bottleneck layer, a ReLU activation function and a batch normalization layer are connected. Finally, the mathematical representation of the Inception module is as follows:

[0131] y = concat([conv(x, W f ) for f in F])

[0132] z = conv(y, W b )

[0133] y out = Relu(batchnorm(z))

[0134] where W f and W b represent convolution kernels.

[0135] This embodiment proposes two Inception modules for architecture construction: the normal block and the reduction block. The input and output dimensions of the normal block are the same, and it contains 4 branches; the output dimension of the reduction block is halved, and it contains 2 branches. The proposed search operation pool includes 3×1 convolution, 5×1 convolution, 7×1 convolution, 9×1 convolution, 11×1 convolution, 3×1 max pooling, 5×1 max pooling, 3×1 average pooling, 5×1 average pooling, and skip connection. Each branch can be one of these ten operations. In both the search stage and the training stage, Inception modules are stacked to form the construction of the network architecture.

[0136] Step S3, model search stage. The purpose of the search stage is to obtain the optimal architecture building blocks and fusion weights.

[0137] In this embodiment, three independent branches are set up. Each branch is responsible for processing feature inputs from different perspectives respectively. Its specific structure is as Figure 5 shown. Within each branch, first, features are obtained through a feature extractor, then feature vectors are generated through a fully connected layer, and then the output is transformed into a probability distribution through a softmax layer. Then, different features are fused through a probability weight fusion method to obtain the final fused probability distribution. Next, the loss is calculated. During the training process, the architecture parameters and fusion weight parameters are continuously optimized.

[0138] Based on the DARTS method, the network architecture most suitable for the input features of this branch is dynamically optimized to ensure that the diversity and specificity of features from different perspectives can be fully explored and utilized. This design can effectively balance the advantages of features from multiple perspectives and enhance the expression ability and domain adaptation ability of the model.

[0139] Specifically, the model is set as a three-branch structure. Branch 1 processes the original signal, Branch 2 processes the discrete wavelet transform features, and Branch 3 processes the short-time Fourier transform features. Each branch searches for the network architecture most suitable for the features of this branch in the Inception search space based on the neural architecture search method DARTS. The results of the three branches are fused at the decision level through dynamic probability weight fusion. A joint loss function L total is designed to guide model optimization, including cross-entropy loss and domain adaptation loss based on maximum mean discrepancy.

[0140] The optimization update of the architecture parameters β and the model parameters ω by DARTS is based on bilevel optimization. Bilevel optimization is divided into upper-level optimization and lower-level optimization. Upper-level optimization is based on the validation set loss L val , calculates the gradient through backpropagation and uses gradient descent for optimization. Lower-level optimization is based on the training set loss L train , and obtains the optimal network weights by minimizing the training error:

[0141]

[0142] After the search is completed, the optimal discrete operation between each node is selected according to the maximum weight of the architecture parameters to generate the final network architecture.

[0143]

[0144] A dynamic probability weight fusion method is proposed to perform weighted fusion on the prediction results of the three branches. The weight assignment is dynamically adjusted according to the contribution degree of each branch to the final prediction result, so as to achieve the efficient fusion of features from different perspectives. If the weight is ω p , then its optimization process is

[0145] A joint loss function is designed to guide the model optimization, including cross-entropy loss and domain adaptation loss based on maximum mean discrepancy. After the outputs of the three branches are fused by probability weights, the cross-entropy loss is calculated:

[0146]

[0147] where N is the number of samples, y n,i is the actual label of the i-th class of the n-th sample, and p n,i is the probability that the n-th sample is predicted as the i-th class.

[0148] The features output by the three branches through their respective feature extractors are concatenated by channels, and the domain adaptation loss between the source domain and the target domain is calculated. This loss can optimize the model in the direction of stronger domain adaptation ability. The calculation method of the domain adaptation loss based on maximum mean discrepancy is as follows:

[0149]

[0150] where m s and m t are the total number of samples in the source domain and the target domain respectively. φ(·) maps each sample to the Hilbert space related to the kernel .

[0151] The final joint loss function is:

[0152] L total = L clf + λL MMD

[0153] where the λ parameter reflects the influence degree of the domain adaptation loss based on maximum mean discrepancy. In the experiment, λ is set to 0.1. In the search stage, the model is trained until convergence to obtain the optimal architecture and optimal probability fusion weight of each branch.

[0154] Step S4: Domain adaptation stage. After the search stage in Step S3, the optimal architecture building blocks of the three branches and the probability fusion weights most suitable for the current task are obtained. Stack the searched normal blocks and reduction blocks 8 times to build a complete network. Place the reduction blocks in the odd layers and the normal blocks in the even layers. The output of each previous module serves as the input of the next module. Train from scratch using the labeled data in the source domain and the unlabeled data in the target domain, and validate on the target domain validation set.

[0155] In the training stage, a model is built based on the optimal Inception architecture building blocks obtained in the search stage according to the number of proxy layers, trained from scratch, and probability weight fusion is performed using the optimal weights to verify the superiority of the architecture and fusion weights obtained in the search stage.

[0156] Test cases

[0157] Compare the method of the present invention with an individual recognition model: a fixed architecture Inception variant network, and five state-of-the-art domain adaptation methods: Maximum Mean Discrepancy (MMD), Domain-Adversarial Neural Network (DANN), Margin Discrepancy Difference (MDD), Joint Adaptation Network (JAN), Domain Adaptation Neural Architecture Search Network (DA-DARTS) through experiments. Use two NVIDIA RTX 2080Ti to train the model, and conduct experiments based on PyTorch version 2.4.0 and CUDA version 12.4. Conduct experiments on four source domains, each corresponding to three different target domains. The experimental results are shown in Table 1, where the results are the accuracy of the target domain test set.

[0158] Table 1

[0159]

[0160] In Table 1 above, compared with the Inception variant network, MMD, DANN, MDD, JAN, and DA-DARTS, the accuracy of SEI-DARTS on the target domain test set has increased by 19.84%, 13.32%, 15.96%, 13.86%, 13.97%, and 3.16% respectively. Therefore, it can be concluded from Table 1 that the method of this case effectively improves the performance of the model in the target domain compared with other baseline methods, verifying the effectiveness of the method. And it has been verified on different source domains and target domains, confirming the generality of the method of the present invention.

[0161] In summary, the Multi-View Representation Fusion used in the method of the present invention constructs a more comprehensive and accurate data representation by integrating the feature information of multiple perspectives, and is applicable to the analysis and recognition of radiation source signals in complex electromagnetic environments. The information of a single perspective is insufficient to accurately represent the characteristics of the radiation source, while multi-perspective data can integrate the advantages of different features, improve the recognition accuracy and system robustness. Even if some perspectives are interfered, the system can still rely on other perspectives to achieve accurate recognition. Combining with the gradient-based Neural Architecture Search (NAS) method, the present invention transforms the discrete network space into a continuous and differentiable optimization problem, and significantly reduces the search time and resource consumption by using gradient descent. By dynamically exploring the network architecture adapted to multi-perspective features and introducing the domain adaptation loss based on the maximum mean discrepancy to enhance the domain adaptation ability, it not only avoids the high cost of manually designing the network, but also improves the individual recognition performance and flexibility of the model in complex scenarios, and realizes an efficient and robust solution.

[0162] It should be noted that the above content only illustrates the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. For those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and retouches can be made, and these improvements and retouches all fall within the protection scope of the claims of the present invention.

Claims

1. A cross-domain radiation source individual recognition method based on multi-view neural architecture search, characterized in that , including the following steps: S1. Multi-perspective data preparation: Collect radio signals, perform pulse detection by the sliding window method after preprocessing, and divide samples; use different feature extraction methods for radio signal samples to obtain different features; the feature extraction methods at least include discrete wavelet transform to obtain the decomposition and reconstruction features of the signal, and short-time Fourier transform to obtain the time-frequency domain features of the signal; S2. Design the search space based on the Inception structure: The search space at least includes Inception modules. The Inception module includes a normal block and a reduction block. The normal block contains 4 branches, and the number of output channels is the same as the number of input channels; the reduction block contains 2 branches, and the number of output channels is halved; the search space contains a search operation pool, and the selectable operations are not unique; each branch of the Inception module will select an operation from the search operation pool, and the outputs of each branch are concatenated along the channel dimension to capture local and global features simultaneously; S3. Model search: Based on the gradient-optimized neural architecture search method, set multiple branches to input different features respectively, search the feature extraction networks of each feature respectively, optimize the network architecture through gradient descent, and obtain the optimal architecture building blocks; set the probability fusion weights, and dynamically update the probability fusion weights during the search process to achieve the integration of multi-perspective information; S4. Domain adaptation: According to the optimal architecture building blocks obtained by searching in step S3, stack the searched normal blocks and reduction blocks 8 times to build a complete network. Place the reduction blocks on the odd layers and the normal blocks on the even layers. The output of each previous module is used as the input of the next module. Build a complete network through the labeled data in the source domain and the unlabeled data in the target domain, and perform domain adaptation from scratch.

2. The cross-domain radiation source individual recognition method based on multi-view neural architecture search according to claim 1, characterized in that: The discrete wavelet transform in step S1 is implemented based on the continuous wavelet transform, and the continuous wavelet transform is defined as: Among them, WT x (a, b) represents the projection of the signal x(t) on the wavelet basis function ψ(t). The discrete wavelet transform discretizes the scale factor a and the translation factor b to obtain a set of discrete wavelet functions: The specific calculation of the discrete wavelet transform is as follows: where represents the wavelet coefficient obtained from the signal x(t) at the scale and the translation kb0, and ψ(·) is the wavelet basis function.

3. The cross-domain radiation source individual recognition method based on multi-view neural architecture search according to claim 1, characterized in that: The short-time Fourier transform in step S1 analyzes the time-frequency domain features of the signal, specifically including the following steps: S11: Select the Hamming window as the window function, and the window function w(t) can be defined as: where T w is the total window length; S12: Divide the original signal into overlapping or non-overlapping short time periods, and perform weighted processing through the window function: x w (t) = x(t) · w(t) S13: Apply the Fourier transform to the weighted signal in each time period, convert the signal from the time domain to the frequency domain, and calculate the spectral information of each time period: where \(X(\omega, t)\) is the frequency component of the signal at time \(t\) and frequency \(\omega\) within the time window from \(t_0\) to \(t_0 + T\). w \(j\) is the scale parameter, and \(e\) -jωt is the Fourier transform kernel function; \(x\) w (t) is the signal obtained by windowing the original signal \(x(t)\) with the window function \(w(t)\), and \(x\) w (t)=x(t)·w(t - t0).

4. The cross-domain radiation source individual recognition method based on multi-view neural architecture search according to claim 1, wherein: The mathematical representation of the Inception module in step S2 is specifically as follows: y = concat([conv(x, W f ) for f in F]) z = conv(y, W b ) y out = Relu(batchnorm(z)) Among them, conv(·) represents the convolution operation, and W f and W b represent the convolution kernels, batchnorm(·) represents batch normalization, Relu(·) is the activation function, F represents the search operation pool. Each operation f in it is traversed and the convolution operation is performed on it. The obtained feature representations are concatenated by channels through concat(·), 1×1 convolution is calculated, and then the module output is obtained through batch normalization and the activation function.

5. The cross-domain radiation source individual recognition method based on multi-view neural architecture search according to claim 4, wherein: In step S2, the search operation pool includes 3×1 convolution, 5×1 convolution, 7×1 convolution, 9×1 convolution, 11×1 convolution, 3×1 max pooling, 5×1 max pooling, 3×1 average pooling, 5×1 average pooling, and skip connection. The optional operation placed in each branch of the Inception module is any one of the above ten operations.

6. The cross-domain radiation source individual recognition method based on multi-view neural architecture search according to claim 1, characterized in that: In the neural architecture search method based on gradient optimization in step S3, it includes double-layer optimization. The double-layer optimization is divided into upper-layer optimization and lower-layer optimization. The architecture parameter α is used as the upper-layer variable, and based on the loss L on the validation set val , calculate the gradient through backpropagation and optimize it using gradient descent; the network weight w is used as the lower-layer variable, and based on the loss L on the training set train , obtain the optimal network weight by minimizing the training error; the specific double-layer optimization is as follows: After the search is completed, according to the maximum weight of the architecture parameters, select the optimal discrete operation between each node to generate the final network architecture: Among them, is the architecture parameter of all optional operations between node i and node j, and the finally selected operation o is obtained by taking argmax (i,j) .

7. The cross-domain radiation source individual recognition method based on multi-view neural architecture search according to claim 6, wherein: The class probability classification after dynamically updating the fusion weights during the search process in step S3 is specifically: P = γ1·P1 + γ2·P2 + … + γ N ·P N Among them, γ1, …, γ N are the set dynamic fusion weights, and P i is the probability distribution obtained by passing the input data of each branch through its respective feature extractor, fully connected layer, and softmax layer.

8. The cross-domain radiation source individual recognition method based on multi-view neural architecture search according to claim 1, characterized in that: The loss function in the field adaptation of step S4 includes cross-entropy loss and domain adaptation loss based on maximum mean discrepancy. The cross-entropy loss is specifically: where N is the number of samples, C is the total number of categories in the classification task, and y n,i is the actual label of the n-th sample for the i-th category, and p n,i is the probability that the n-th sample is predicted to be the i-th category; The domain adaptation loss is specifically: Among them is the source domain feature, is the target domain feature, m S and m T are the total numbers of samples in the source domain and the target domain respectively. φ(·) maps each sample to the Hilbert space related to the kernel space; The final combined loss function is: L total = L clf + λL MMD where λ is a parameter reflecting the influence degree of the domain adaptation loss based on maximum mean discrepancy.

Images