Radio frequency device identification method and apparatus

Abstract

The application discloses a radio frequency equipment identity recognition method based on a complex-valued convolutional neural network and a Conformer encoder hybrid architecture. The method sets feature extraction and classification as a unified end-to-end task, extracts local time-frequency features of I / Q signals through a complex-valued convolutional neural network, captures global time sequence dependency through a Conformer encoder, adopts a double-assistant classifier deep supervision training method, a residual fusion mechanism, and a P x K sampling strategy-based measurement learning strategy combined with batch-hard triplet mining, can effectively process complex I / Q signals, takes into account local feature extraction and global dependency modeling, and significantly improves the accuracy and stability of radio frequency equipment identity recognition.

Description

Technical Field

[0001] This invention belongs to the technical field of wireless communication security and signal processing, and particularly relates to a method and apparatus for identifying radio frequency devices, which is used to achieve high-precision extraction and identification of radio frequency fingerprints based on complex numerical neural networks and Conformer architecture. Background Technology

[0002] Radio frequency (RF) fingerprinting is a technology that uses inherent hardware characteristics in the signals emitted by wireless devices to uniquely identify them. Each wireless transmitter, due to minute differences in the manufacturing process, leaves a unique fingerprint in its emitted radio frequency signals. These fingerprint features are uncopyable, thus enabling highly secure device authentication.

[0003] Radio frequency fingerprint recognition faces three core challenges: First, traditional real-valued convolutional neural networks (CNNs) typically process I / Q complex signals separately or simply concatenate the real and imaginary parts, failing to effectively model the phase information and correlation between the real and imaginary parts of complex signals, resulting in insufficient feature extraction. Second, existing CNN methods are limited by local receptive fields and lack the ability to model long-distance temporal dependencies, making it difficult to capture global temporal features in RF signals. Third, traditional classification loss functions only focus on class boundaries, failing to explicitly optimize intra-class compactness and inter-class separability of the feature space, leading to insufficient generalization ability in scenarios with few samples. Furthermore, deep neural networks suffer from unstable training and severe overfitting, especially with limited datasets.

[0004] To address the aforementioned issues, there is an urgent need for a radio frequency device identification technology solution that can effectively process complex signals, take into account both local and global features, possess good metric learning capabilities, and be stably trained. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for identifying radio frequency (RF) devices, which unifies feature extraction and classification into an end-to-end task to achieve high-precision and high-stability RF device identification.

[0006] To solve the above problems, the technical solution of the present invention is as follows: A method for identifying radio frequency devices, based on a hybrid architecture of complex numerical convolutional neural network and Conformer encoder, includes the following steps: Signal preprocessing steps: Acquire the I / Q signal transmitted by the radio frequency device, combine the in-phase component I(t) and the quadrature component Q(t) into a complex signal, perform segmentation processing on the complex signal, and splice the real and imaginary parts of the complex signal along the channel dimension to form a real tensor as the network input, and perform normalization processing. Feature extraction and global modeling steps: Construct a hybrid architecture of complex numerical convolutional neural network and Conformer encoder. The input tensor obtained from the signal preprocessing step is sequentially passed through the feature extraction module of complex numerical convolutional neural network to extract local time-frequency features, and through the global modeling module of Conformer encoder to capture global temporal dependencies. Feature fusion and classification steps: The output features of the complex numerical convolutional neural network and the output features of the Conformer encoder are fused through the residual fusion layer, and the classification result is output through the deep classification head; Model training steps: A training batch is constructed using a P×K sampling strategy, and the hybrid architecture is trained based on a joint loss function, which integrates classification loss and metric learning loss; Reasoning and recognition steps: Use the trained model to reason about the signal to be recognized and output the device identity recognition result.

[0007] According to an embodiment of the present invention, the complex numerical convolutional neural network feature extraction module includes 9 complex convolutional blocks, each of which sequentially performs complex convolution, ReLU activation and complex batch normalization operations; The complex convolution operation follows the complex multiplication rule. Let the input x = xr + jxi, and the convolution kernel W = Wr + jWi, then the formula for calculating the output y = yr + jyi is: yr = Wr *xr - Wi *xi yi = Wr * xi + Wi * xr Where * represents convolution operation.

[0008] According to an embodiment of the present invention, the complex numerical convolutional neural network employs a complex Kaiming initialization method: the magnitude follows a Rayleigh distribution, the phase follows a uniform distribution in [0, 2π], and the real and imaginary weights are obtained through polar coordinate transformation. in Z follows a standard normal distribution. , where fanin is the product of the number of input channels and the kernel size; The complex batch normalization employs 2×2 covariance matrix whitening, where the covariance matrix is ​​defined as: in, For the real part variance, For the imaginary part of the variance, is the covariance of the real and imaginary parts; the whitening transformation matrix is ​​the inverse square root of the covariance matrix.

[0009] According to one embodiment of the present invention, the global modeling module of the Conformer encoder adopts a two-layer Macaron structure, each layer sequentially including a first feedforward network module, a multi-head self-attention module, a convolution module, and a second feedforward network module; The multi-head self-attention module adopts a 4-head attention mechanism, with each head having a dimension of 32, and introduces relative position encoding. The attention calculation formula is as follows: Where Q, K, and V are the query, key, and value matrices, respectively; Pos is the relative position encoding matrix; d is the dimension per head; and u and v are learnable position bias parameters. The convolution module sequentially includes layer normalization, point convolution, gated linear unit, depthwise separable convolution, batch normalization, Swish activation, and point convolution, and uses residual connections.

[0010] According to an embodiment of the present invention, the residual fusion layer uses residual connections for feature fusion, and the fusion formula is as follows: in For the output features of a complex numerical convolutional neural network, The output features of the Conformer encoder, where scale is the scaling factor; The deep classification head consists of three layers: the first layer reduces the fused features from 9600 dimensions to 512 dimensions, the second layer reduces the fused features from 512 dimensions to 256 dimensions and uses the output of this layer as the embedded features for metric learning, and the third layer maps from 256 dimensions to the number of categories C.

[0011] According to one embodiment of the present invention, the hybrid architecture further includes a dual-auxiliary classifier deep supervised training mechanism, wherein a first auxiliary classifier is set at the output position of the 9th layer of the complex numerical convolutional neural network, and a second auxiliary classifier is set at the output position of the 2nd layer of the Conformer encoder.

[0012] According to an embodiment of the present invention, the joint loss function is defined as: in The cross-entropy loss of the main classifier, For triple loss, Loss at the center and The cross-entropy loss of the two auxiliary classifiers are respectively. , , These are the weighting coefficients for each loss.

[0013] According to one embodiment of the present invention, the triplet loss employs a batch hard triplet mining strategy, whereby for each anchor sample a, the farthest positive sample in the same class and the closest negative sample in different classes are selected: The triplet loss is defined as: Where A is the set of anchor points, margin is the boundary parameter, and d(·,·) is the Euclidean distance; The central loss is defined as: in For category The feature centers are N, where N is the number of samples in the batch.

[0014] According to one embodiment of the present invention, the model training uses the AdamW optimizer, with a base learning rate and weight decay configured; a differentiated learning rate strategy is adopted: the complex numerical convolutional neural network and the classification head use the base learning rate, and the Conformer encoder uses 1.5 times the base learning rate; The learning rate scheduling adopts a preheating plus cosine annealing strategy. The first n training rounds are linearly preheated, and then the learning rate decays according to the cosine function.

[0015] A radio frequency device identification device, based on a hybrid architecture of complex numerical convolutional neural network and Conformer encoder, includes: The signal preprocessing module is configured to acquire the I / Q signals transmitted by the radio frequency device, combine the in-phase component I(t) and the quadrature component Q(t) into a complex signal, perform segmentation processing on the complex signal, and splice the real and imaginary parts of the complex signal along the channel dimension to form a real tensor as the network input, and perform normalization processing. The feature extraction and global modeling module is configured to construct a hybrid architecture of complex numerical convolutional neural network and Conformer encoder. The input tensor obtained from the signal preprocessing step is sequentially passed through the complex numerical convolutional neural network feature extraction module to extract local time-frequency features and through the Conformer encoder global modeling module to capture global temporal dependencies. The feature fusion and classification module is configured to fuse the output features of the complex numerical convolutional neural network and the output features of the Conformer encoder through a residual fusion layer, and output the classification result through the deep classification head; The model training module is configured to construct training batches using a P×K sampling strategy and train the hybrid architecture based on a joint loss function, which integrates classification loss and metric learning loss. The reasoning and recognition module is configured to use the trained model to reason about the signal to be recognized and output the device identity recognition result.

[0016] Because the present invention adopts the above technical solution, it has the following advantages and positive effects compared with the prior art: The radio frequency device identification method in one embodiment of the present invention, by constructing an end-to-end unified architecture, combines the local time-frequency feature extraction capability of complex numerical convolutional neural networks with the global temporal dependency modeling capability of Conformer encoders, effectively solving the problems of phase information loss and insufficient long-distance temporal feature capture in traditional methods when processing I / Q complex signals. At the same time, by adopting a P×K sampling strategy and a joint loss function that fuses classification loss and metric learning, the intra-class compactness and inter-class separability of the feature space are explicitly optimized, significantly improving the generalization ability and recognition stability of the model in low-sample scenarios. Furthermore, the training stability of deep networks is enhanced through a residual fusion mechanism, thereby achieving high accuracy and high reliability in radio frequency device identification. Attached Figure Description

[0017] Figure 1 This is a flowchart of a radio frequency device identification method according to an embodiment of the present invention; Figure 2 This is a block diagram of a hybrid architecture of complex numerical convolutional neural network and Conformer encoder in one embodiment of the present invention; Figure 3 This is a performance comparison chart between the radio frequency device identification method in one embodiment of the present invention and a baseline method; Figures 4-6 This is a visualization result of t-SNE for 10, 20, and 30 feature spaces in one embodiment of the present invention. Detailed Implementation

[0018] The following detailed description of the radio frequency device identification method and apparatus proposed in this invention, in conjunction with the accompanying drawings and specific embodiments, provides further insight into this invention.

[0019] This embodiment addresses the core problems in radio frequency fingerprint recognition, such as insufficient complex signal processing, weak global feature modeling capability, poor feature space separability, and unstable training. It provides a radio frequency device identification method based on a hybrid architecture of complex numerical convolutional neural network and Conformer encoder, which unifies feature extraction and classification into an end-to-end task, achieving high-precision and high-stability radio frequency device identification.

[0020] Please refer to Figure 1 The radio frequency device identification method, based on a hybrid architecture of complex numerical convolutional neural network and Conformer encoder, includes the following steps: S1: Signal preprocessing steps: acquire the I / Q signal transmitted by the radio frequency device, combine the in-phase component I(t) and the quadrature component Q(t) into a complex signal, perform segmentation processing on the complex signal, and splice the real and imaginary parts of the complex signal along the channel dimension to form a real tensor as the network input, and perform normalization processing. S2: Feature extraction and global modeling steps: Construct a hybrid architecture of complex numerical convolutional neural network and Conformer encoder. The input tensor obtained from the signal preprocessing step is sequentially passed through the feature extraction module of complex numerical convolutional neural network to extract local time-frequency features and through the global modeling module of Conformer encoder to capture global temporal dependencies. S3: Feature fusion and classification steps: The output features of the complex numerical convolutional neural network and the output features of the Conformer encoder are fused through the residual fusion layer, and the classification result is output through the deep classification head; S4: Model training steps: P×K sampling strategy is used to construct training batches, and the hybrid architecture is trained based on the joint loss function, which integrates classification loss and metric learning loss; S5: Reasoning and Recognition Steps: Use the trained model to reason about the signal to be recognized and output the device identity recognition result.

[0021] For details, please refer to Figure 2 In step S1, a software-defined radio device is used as a receiver to acquire the signal transmitted by the target radio frequency device. The receiver performs down-conversion processing on the radio frequency signal to obtain the in-phase component of the baseband. and orthogonal components .Will As the actual part, As the imaginary part, they combine to form a complex baseband signal. The acquired complex signal is segmented, with each sample containing... A series of consecutive sampling points, in sequence form: The real and imaginary parts of the complex signal are then concatenated along the channel dimension to form an input tensor of dimension [B, 4800, 2]. The dataset used in this embodiment contains 90 different categories of RF devices, with a total of approximately 20,566 samples. The training and validation sets are divided in a 9:1 ratio, resulting in 18,509 training samples and 2,057 validation samples. Normalization parameters are calculated only from the training set. and Normalize all data:

[0022] In step S2, the complex numerical convolutional neural network contains 9 layers of complex convolutional blocks. Each layer performs complex convolution, ReLU activation, and complex batch normalization operations in the order of convolution-activation-normalization. The complex convolution operation uses two independent sets of convolution kernels to represent the real parts of the complex weights. and the virtual part The output is: The weights of the complex convolutional layer are initialized using the complex Kaiming method: the magnitude follows a Rayleigh distribution, and the phase follows a uniform distribution in [0, 2π]. The real and imaginary weights are obtained through polar coordinate transformation. in Z follows a standard normal distribution. , This is the product of the number of input channels and the kernel size. Complex batch normalization uses 2×2 covariance matrix whitening, where the covariance matrix is ​​defined as: in For the real part variance, For the imaginary part of the variance, The real and imaginary parts are covariances; the whitening transformation matrix is ​​the square root of the inverse of the covariance matrix; the affine transformation parameters are... The diagonal elements are initialized to Off-diagonal elements are initialized to 0, and the numerical stability constant is... The momentum parameter is 0.9.

[0023] Network configuration: The first layer of complex convolutional block: the number of input channels is 1, the number of output channels is 64, the kernel size is 3, the padding is 1, and max pooling is performed (the pooling kernel size is 2, the stride is 2). Complex convolutional blocks in layers 2-6: 64 input channels, 64 output channels, kernel size of 3, padding of 1, and max pooling is performed. Complex convolutional blocks in layers 7-9: 64 input channels, 64 output channels, kernel size of 3, padding of 1, no pooling operation; Max pooling employs a separation strategy, performing max pooling on the real and imaginary parts separately before concatenating them. After 6 pooling iterations, the sequence length decreases from 4800 to 75, and the output feature dimension of the complex numerical convolutional neural network is [B, 75, 128], where 128 represents the concatenation of the 64-channel real part and the 64-channel imaginary part.

[0024] The Conformer encoder adds a normalization adaptation layer after the output of the complex numerical convolutional neural network, using a two-layer Macaron structure, each layer containing: The first feedforward network module: After layer normalization, it undergoes two linear transformations with an expansion factor of 4, expanding the dimension from 128 to 512 and then reducing it back to 128. The Swish activation function is used in between, and the Swish activation function is defined as follows: The output is multiplied by a scaling factor of 0.5 and then connected using residuals. Multi-head self-attention module: Employs a 4-head attention mechanism, with each head having a dimension of 32, and introduces relative position encoding; attention calculation uses learnable positional bias parameters. and The calculation formula is: Where Q, K, and V are the query, key, and value matrices, respectively, Pos is the relative position encoding matrix, and d is the dimension per head; The convolution module consists of layer normalization, pointwise convolution (expanding the number of channels from 128 to 512), gated linear units (halving the number of channels to 256), depthwise separable convolution (kernel size of 15), batch normalization, Swish activation, and pointwise convolution (restored to 128 channels), using residual connections; the gated linear units are defined as follows: , ⊙ represents element-wise multiplication (Hadamard product); Second feedforward network module: The structure is the same as the first feedforward network module; Final layer normalization: Perform layer normalization on the output of the above modules; A random depth regularization strategy is adopted, and the probability of dropping increases linearly with the depth of the layer, with 0 for the first layer and 0.1 for the second layer.

[0025] The residual fusion layer uses residual connections for feature fusion. Let the output of the complex numerical convolutional neural network be... The Conformer encoder output is The fusion formula is: in The scaling factor is set to 1.0; after layer normalization, the fused features are flattened into a 9600-dimensional (75×128) feature vector. The deep classification head consists of three layers: Layer 1: Linear layer (9600 to 512 dimensions), batch normalization, ReLU activation, dropout layer (dropout probability 0.3); Layer 2: Linear layer (512 to 256 dimensions), batch normalization, ReLU activation, dropout layer (dropout probability 0.2), the 256-dimensional output of this layer is used as embedded features for metric learning; Layer 3: Linear layer (256 to C dimensions), where C is the number of categories.

[0026] A first auxiliary classifier is set at the output position of the 9th layer of the complex numerical convolutional neural network, and a second auxiliary classifier is set at the output position of the 2nd layer of the Conformer encoder; Each auxiliary classifier has the following structure: adaptive average pooling layer (compresses the sequence dimension to 1), linear layer (128 to 256 dimensions), batch normalization, ReLU activation, dropout layer (dropout probability 0.3), and linear classification layer (256 to C dimensions). The auxiliary classifier loss uses the cross-entropy loss function and employs label smoothing technology with a smoothing coefficient of 0.1.

[0027] A dual-auxiliary classifier deep supervised training mechanism is adopted. By introducing auxiliary classifiers into the deep layer (layer 9) of the complex numerical convolutional neural network and the output of the Conformer encoder (layer 2), additional gradient feedback paths are provided for the intermediate layers of the network. This effectively alleviates the gradient vanishing problem during the training of deep neural networks and accelerates the network convergence speed. At the same time, the supervision signal of the auxiliary classifier can guide the intermediate layers to learn more discriminative feature representations, enhance the robustness of feature extraction, and reduce the risk of model overfitting when combined with label smoothing technology. Thus, the training stability and final recognition accuracy of radio frequency device identification are significantly improved on a limited-scale dataset.

[0028] In step S3, a P×K sampling strategy is used to construct training batches, where P = 8, K = 4, and batch size = 32. This strategy ensures that each batch contains 448 pairs of out-of-class samples and 48 pairs of similar samples, providing sufficient positive and negative sample combinations for triplet mining.

[0029] In step S4, the embedded features are first L2 normalized to map them to a unit hypersphere: A batch hard triplet mining strategy is adopted. For each anchor sample a, the positive sample that is furthest away in the same class and the negative sample that is closest in different classes are selected. in For Euclidean distance, Let y be the embedding feature of the sample, and y be the class label; The triplet loss is defined as: in For the set of anchor points, The boundary parameter is set to 0.3; The central loss is defined as: in For category Feature center, This represents the batch sample size; to prevent training instability, the central loss is pruned, with a maximum limit of 100. The joint loss function is defined as: in The cross-entropy loss of the main classifier (using label smoothing with a smoothing coefficient of 0.1). and The cross-entropy loss of the two auxiliary classifiers are respectively. = 0.01, = 0.1, = 0.15; The model was trained using the AdamW optimizer with a base learning rate of 3×10⁻⁶. -4 The weight decays to 1×10 -4 A differentiated learning rate strategy is adopted: the complex numerical convolutional neural network and the classification head use the base learning rate, while the Conformer encoder uses 1.5 times the base learning rate. The learning rate scheduling adopts a preheating plus cosine annealing strategy. The first 15 training rounds are linearly preheated, and then the learning rate is decayed to close to 0 according to the cosine function. The gradient clipping threshold was set to 1.0; the total number of training epochs was set to 150, and an early stopping mechanism with a patience value of 30 was used.

[0030] In step S5, the trained model is used for inference to obtain the device identification result. The optimal model parameters and normalization parameters are loaded, the signal to be identified is segmented and normalized, and then fed into the model sequentially through a complex numerical convolutional neural network, a Conformer encoder, a residual fusion layer, and a deep classification head. The output logits are normalized using Softmax to obtain the predicted probability distribution for each category. The category with the highest probability is selected as the result of radio frequency device identification: .

[0031] To further verify the effectiveness of this method, its recognition performance and feature distribution are analyzed. For example... Figure 3As shown, under the same dataset and experimental conditions, our method is compared with commonly used baseline methods such as K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest (RF), Multilayer Perceptron (MLP), and Lightweight Gradient Boosting Machine (LGBM). KNN and SVM are two classic traditional machine learning classification algorithms that primarily rely on manually designed features to complete signal pattern matching; RF and LGBM belong to the category of ensemble learning, effectively improving the model's generalization ability through the combined application of multiple decision trees, and are widely used in various classification tasks; MLP, as a basic deep learning model, achieves effective feature mapping through fully connected layers. Figure 3 As shown, our proposed method significantly outperforms the aforementioned baseline methods in the task of identifying radio frequency (RF) devices. This fully demonstrates that the hybrid architecture combining complex numerical convolutional neural networks and a Conformer encoder can more accurately capture the key features of RF signals, thereby effectively improving the recognition performance. The superior performance of our proposed method in RF device identification demonstrates that the hybrid architecture of complex numerical convolutional neural networks and a Conformer encoder can effectively improve the recognition results.

[0032] like Figures 4-6 As shown, the embedded features extracted by this method are visualized using t-SNE. In scenarios with 10, 20, and 30 device classes, samples of different classes show good clustering effects and separation trends, indicating that this method has good intra-class compactness and inter-class separability.

[0033] Based on the same concept, this application also provides a radio frequency device identification device, based on a hybrid architecture of complex numerical convolutional neural network and Conformer encoder, including: The signal preprocessing module is configured to acquire the I / Q signals transmitted by the radio frequency device, combine the in-phase component I(t) and the quadrature component Q(t) into a complex signal, perform segmentation processing on the complex signal, and splice the real and imaginary parts of the complex signal along the channel dimension to form a real tensor as the network input, and perform normalization processing. The feature extraction and global modeling module is configured to construct a hybrid architecture of complex numerical convolutional neural network and Conformer encoder. The input tensor obtained from the signal preprocessing step is sequentially passed through the complex numerical convolutional neural network feature extraction module to extract local time-frequency features and through the Conformer encoder global modeling module to capture global temporal dependencies. The feature fusion and classification module is configured to fuse the output features of the complex numerical convolutional neural network and the output features of the Conformer encoder through a residual fusion layer, and output the classification result through the deep classification head; The model training module is configured to construct training batches using a P×K sampling strategy and train the hybrid architecture based on a joint loss function, which integrates classification loss and metric learning loss. The reasoning and recognition module is configured to use the trained model to reason about the signal to be recognized and output the device identity recognition result.

[0034] This device is used to implement the above-mentioned radio frequency device identification method, and its implementation method is similar, so it will not be described in detail here.

[0035] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments. Even if various changes are made to the present invention, if these changes fall within the scope of the claims of the present invention and their equivalents, they shall still fall within the protection scope of the present invention.

Claims

1. A method for identifying radio frequency devices, based on a hybrid architecture of complex numerical convolutional neural network and Conformer encoder, characterized in that, Includes the following steps: Signal preprocessing steps: Acquire the I / Q signal transmitted by the radio frequency device, combine the in-phase component I(t) and the quadrature component Q(t) into a complex signal, perform segmentation processing on the complex signal, and splice the real and imaginary parts of the complex signal along the channel dimension to form a real tensor as the network input, and perform normalization processing. Feature extraction and global modeling steps: Construct a hybrid architecture of complex numerical convolutional neural network and Conformer encoder. The input tensor obtained from the signal preprocessing step is sequentially passed through the feature extraction module of complex numerical convolutional neural network to extract local time-frequency features, and through the global modeling module of Conformer encoder to capture global temporal dependencies. Feature fusion and classification steps: The output features of the complex numerical convolutional neural network and the output features of the Conformer encoder are fused through the residual fusion layer, and the classification result is output through the deep classification head; Model training steps: A training batch is constructed using a P×K sampling strategy, and the hybrid architecture is trained based on a joint loss function, which integrates classification loss and metric learning loss; Reasoning and recognition steps: Use the trained model to reason about the signal to be recognized and output the device identity recognition result.

2. The radio frequency device identification method as described in claim 1, characterized in that, The complex numerical convolutional neural network feature extraction module contains 9 layers of complex convolutional blocks, each of which performs complex convolution, ReLU activation and complex batch normalization operations in sequence. The complex convolution operation follows the complex multiplication rule. Let the input x = xr + jxi, and the convolution kernel W = Wr + jWi, then the formula for calculating the output y = yr + jyi is: yr = Wr *xr - Wi *xi yi = Wr * xi + Wi * xr Where * represents convolution operation.

3. The radio frequency device identification method as described in claim 2, characterized in that, The complex numerical convolutional neural network employs the complex Kaiming initialization method: the magnitude follows a Rayleigh distribution, and the phase follows a uniform distribution in [0, 2π]. The real and imaginary weights are obtained through polar coordinate transformation. in Z follows a standard normal distribution. , where fanin is the product of the number of input channels and the kernel size; The complex batch normalization employs 2×2 covariance matrix whitening, where the covariance matrix is ​​defined as: in, For the real part variance, For the imaginary part of the variance, is the covariance of the real and imaginary parts; the whitening transformation matrix is ​​the inverse square root of the covariance matrix.

4. The radio frequency device identification method as described in claim 1, characterized in that, The global modeling module of the Conformer encoder adopts a two-layer Macaron structure, with each layer containing a first feedforward network module, a multi-head self-attention module, a convolution module, and a second feedforward network module in sequence. The multi-head self-attention module adopts a 4-head attention mechanism, with each head having a dimension of 32, and introduces relative position encoding. The attention calculation formula is as follows: Where Q, K, and V are the query, key, and value matrices, respectively; Pos is the relative position encoding matrix; d is the dimension per head; and u and v are learnable position bias parameters. The convolution module sequentially includes layer normalization, point convolution, gated linear unit, depthwise separable convolution, batch normalization, Swish activation, and point convolution, and uses residual connections.

5. The radio frequency device identification method as described in claim 1, characterized in that, The residual fusion layer uses residual connections for feature fusion, and the fusion formula is as follows: in For the output features of a complex numerical convolutional neural network, The output features of the Conformer encoder, where scale is the scaling factor; The deep classification head consists of three layers: the first layer reduces the fused features from 9600 dimensions to 512 dimensions, the second layer reduces the fused features from 512 dimensions to 256 dimensions and uses the output of this layer as the embedded features for metric learning, and the third layer maps from 256 dimensions to the number of categories C.

6. The radio frequency device identification method as described in claim 1, characterized in that, The hybrid architecture also includes a dual-aid classifier deep supervised training mechanism, in which a first aid classifier is set at the output position of the 9th layer of the complex numerical convolutional neural network, and a second aid classifier is set at the output position of the 2nd layer of the Conformer encoder.

7. The radio frequency device identification method as described in claim 1, characterized in that, The joint loss function is defined as follows: in The cross-entropy loss of the main classifier, For triple loss, Loss at the center and The cross-entropy loss of the two auxiliary classifiers are respectively. , , These are the weighting coefficients for each loss.

8. The radio frequency device identification method as described in claim 7, characterized in that, The triplet loss employs a batch hard triplet mining strategy. For each anchor sample 'a', the farthest positive sample in the same class and the closest negative sample in different classes are selected. The triplet loss is defined as: Where A is the set of anchor points, margin is the boundary parameter, and d(·,·) is the Euclidean distance; The central loss is defined as: in For category The feature centers are N, where N is the number of samples in the batch.

9. The radio frequency device identification method as described in claim 1, characterized in that, The model training uses the AdamW optimizer with a base learning rate and weight decay configured; a differentiated learning rate strategy is adopted: the complex numerical convolutional neural network and the classification head use the base learning rate, while the Conformer encoder uses 1.5 times the base learning rate; The learning rate scheduling adopts a preheating plus cosine annealing strategy. The first n training rounds are linearly preheated, and then the learning rate decays according to the cosine function.

10. A radio frequency device identification device, based on a hybrid architecture of complex numerical convolutional neural network and Conformer encoder, characterized in that, include: The signal preprocessing module is configured to acquire the I / Q signals transmitted by the radio frequency device, combine the in-phase component I(t) and the quadrature component Q(t) into a complex signal, perform segmentation processing on the complex signal, and splice the real and imaginary parts of the complex signal along the channel dimension to form a real tensor as the network input, and perform normalization processing. The feature extraction and global modeling module is configured to construct a hybrid architecture of complex numerical convolutional neural network and Conformer encoder. The input tensor obtained from the signal preprocessing step is sequentially passed through the complex numerical convolutional neural network feature extraction module to extract local time-frequency features and through the Conformer encoder global modeling module to capture global temporal dependencies. The feature fusion and classification module is configured to fuse the output features of the complex numerical convolutional neural network and the output features of the Conformer encoder through a residual fusion layer, and output the classification result through the deep classification head; The model training module is configured to construct training batches using a P×K sampling strategy and train the hybrid architecture based on a joint loss function, which integrates classification loss and metric learning loss. The reasoning and recognition module is configured to use the trained model to reason about the signal to be recognized and output the device identity recognition result.

Images