C-v2x radio frequency fingerprinting method based on multi-channel robust feature fusion

Abstract

The application discloses a C-V2X radio frequency fingerprint identification method based on multi-channel robust feature fusion, relates to the technical field of radio frequency fingerprint identification, and receives a physical sidelink broadcast channel baseband signal of a transmitter through a receiver, extracts an effective subcarrier sequence, further extracts three kinds of channel robust features respectively, and remodels into a single-channel two-dimensional feature map at the same time, simultaneously performs complex value feature extraction and scale alignment processing, extracts I / Q self features and I / Q coupling features, completes feature scale calibration, obtains three normalized feature representations, further adopts adaptive fusion processing, fuses the three normalized feature representations through weight adaptive distribution, obtains a fused feature representation, finally processes the obtained fused feature representation through final extraction and matching identification operation of the radio frequency fingerprint, and obtains a radio frequency fingerprint identification result, so that the equipment identity recognition precision and operation efficiency in a complex scene are improved.

Description

Technical Field

[0001] This invention relates to the field of radio frequency fingerprint recognition technology, specifically to a C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion. Background Technology

[0002] C-V2X is a core technology for secure communication in vehicle-to-everything (V2X) communication. Its device authentication relies on radio frequency fingerprinting technology. This technology uses unique signal characteristics brought about by differences in transmitter hardware to achieve identification. It usually extracts features based on the effective subcarrier sequence of the Physical Side Link Broadcast Channel (PSBCH). Existing solutions mostly use single features such as channel response and cepstrum to complete fingerprint construction, which can achieve basic identity verification in simple scenarios.

[0003] In real-world vehicle-to-everything (V2X) scenarios, channel fading and multipath interference can significantly reduce the robustness of a single feature. Meanwhile, existing technologies do not fully exploit the inherent and coupled information of the I / Q components, and lack adaptive weight allocation during multi-feature fusion. The dimensional differences between different features can easily lead to information loss, resulting in insufficient recognition accuracy and low computational efficiency in complex scenarios. This makes it difficult to meet the high reliability and low latency authentication requirements of C-V2X, thus necessitating an optimized RF fingerprint recognition scheme that integrates multiple features. Summary of the Invention

[0004] The purpose of this invention is to provide a C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion, which solves the problems existing in the background technology.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion, specifically including the following steps:

[0006] S1, the receiver receives the baseband signal of the physical side link broadcast channel of the transmitter and extracts the effective subcarrier sequence;

[0007] The effective subcarrier sequences include: PSSS, SSSS, and DMRS;

[0008] S2, Based on the extracted effective subcarrier sequence, three types of channel robust features are extracted respectively, and the three types of channel robust features are reshaped into single-channel two-dimensional feature maps respectively.

[0009] The three channel robustness features include: channel equalization feature, nearest-to-nearest-symbol logarithmic difference complex cepstrum feature, and fractional-domain symbol self-ratio feature;

[0010] S3, perform complex value feature extraction and scale alignment processing on the reshaped single-channel two-dimensional feature map, extract the I / Q features and I / Q coupling features, and complete the feature scale calibration to obtain the three-way normalized feature representation;

[0011] S4. Based on the obtained three-way feature representations, an adaptive fusion process is adopted, and the three normalized feature representations are fused through adaptive weight allocation to obtain a fused feature representation.

[0012] S5 processes the obtained fused feature representation through the final extraction and matching recognition operation of the radio frequency fingerprint to obtain the radio frequency fingerprint recognition result.

[0013] Preferably, the step of receiving the baseband signal of the physical side link broadcast channel of the transmitter through the receiver and extracting the effective subcarrier sequence includes the following steps:

[0014] The receiver receives the baseband signal from the physical side link broadcast channel of the transmitter and obtains complete baseband signal data based on the received baseband signal.

[0015] From the complete baseband signal data, the effective subcarrier sequences corresponding to PSSS, SSSS and DMRS are separated and extracted respectively. The effective subcarrier sequences of PSSS and SSSS each contain 62 subcarriers, and the effective subcarrier sequence of DMRS contains 72 subcarriers.

[0016] Preferably, the step of extracting three channel-robust features based on the extracted effective subcarrier sequences, and reshaping the three extracted channel-robust features into single-channel two-dimensional feature maps, includes the following steps:

[0017] S21, perform least squares estimation on the effective subcarrier sequences corresponding to the extracted PSSS, SSSS and DMRS to obtain the channel equalization characteristics;

[0018] S22, the second PSSS and the first DMRS, and the first SSSS and the third DMRS in the physical side link broadcast channel subframe are taken as two sets of closest opposite symbol pairs. At the same time, 62 subcarrier sequences at the center of the first DMRS and the third DMRS are extracted. The channel influence is canceled by logarithmic difference operation and transformed to the cepstral domain by inverse Fourier transform to obtain the closest opposite symbol logarithmic difference complex cepstral features.

[0019] S23, select the effective subcarrier sequences of two PSSS and the first two DMRS in the physical side link broadcast channel, take the average and denoise them respectively, divide the averaged PSSS and DMRS into four subsequences, divide the first half subsequence with the second half subsequence to obtain the symbol self-ratio value, and map it to the fractional domain through inverse Fourier transform and fractional Fourier transform to obtain the fractional domain symbol self-ratio value characteristics.

[0020] S24. Based on the lengths of the I and Q components and the channel robustness features, the height and width of the feature map are set respectively. The extracted channel equalization features, the nearest opposite symbol logarithmic difference complex cepstral features, and the fractional domain symbol self-ratio features are reshaped into single-channel two-dimensional feature maps respectively.

[0021] Preferably, the step of taking the second PSSS and the first DMRS, and the first SSSS and the third DMRS in the physical side link broadcast channel subframe as two sets of closest opposite symbol pairs, and simultaneously extracting 62 subcarrier sequences from the centers of the first DMRS and the third DMRS, canceling the channel influence through logarithmic difference operation, and converting to the cepstral domain through inverse Fourier transform to obtain the closest opposite symbol logarithmic difference complex cepstral features includes the following steps:

[0022] In the physical side link broadcast channel subframe, the second PSSS and the first DMRS, and the first SSSS and the third DMRS are selected as two sets of closest opposite symbol pairs, and 62 subcarrier sequences at the center of the first DMRS and the third DMRS are extracted;

[0023] Logarithmic operations are performed on the sequences corresponding to the two sets of opposite symbol pairs, and then the channel influence is canceled out by pairwise subtraction. The results are then transformed into the cepstrum domain using an inverse Fourier transform to obtain the most recent opposite symbol logarithmic difference complex cepstrum features.

[0024] The expression is as follows:

[0025] ;

[0026] because and Approximately equal and Since they are approximately equal, the above formula can be further expressed as:

[0027] ;

[0028] The inverse Fourier transform is shown below:

[0029] ;

[0030] in The difference between the two most recent pairs of opposite-signed logarithms. These are the channel frequency responses for the second PSSS, the first DMRS, the first SSSS, and the third DMRS, respectively. These are the sequences carrying radio frequency fingerprint information for the second PSSS, the first DMRS, the first SSSS, and the third DMRS, respectively. For the second PSSS, the effective subcarrier sequence, This is the first DMRS subcarrier sequence. This is the first valid subcarrier sequence of SSSS. This is the third DMRS subcarrier sequence. For two most recent log-difference complex cepstral differences with different signs, This is the inverse Fourier transform.

[0031] Preferably, the step of selecting two effective subcarrier sequences of the physical side link broadcast channel (PSSS) and the first two DMRS, averaging and denoising them respectively, dividing the averaged PSSS and DMRS into four subsequences, dividing the first half of the subsequence by the second half to obtain the symbol self-ratio, and mapping it to the fractional domain through inverse Fourier transform and fractional Fourier transform to obtain the fractional domain symbol self-ratio characteristics includes the following steps:

[0032] Select the effective subcarrier sequences of the two PSSS and the first two DMRS in the physical side link broadcast channel, take the average noise reduction, divide them into four subsequences, and divide the first half subsequence with the second half subsequence to obtain the symbol self-ratio value.

[0033] Perform inverse Fourier transform and fractional Fourier transform on the sign self-ratio in sequence, and map it to the fractional domain to obtain the fractional domain sign self-ratio characteristics;

[0034] The expression for the symbolic self-ratio is shown below:

[0035] ;

[0036] The characteristics of the sign-based ratio in the fractional domain are shown below:

[0037] ;

[0038] in The sign-to-value ratios of PSSS and DMRS are given. It is a four-segment subsequence that is equally divided. The fractional sign ratios of PSSS and DMRS are respectively. for Fractional Fourier transform, This is the transformation kernel function.

[0039] Preferably, the process of performing complex-valued feature extraction and scale alignment on the reconstructed single-channel two-dimensional feature map, extracting I / Q features and I / Q coupling features, and calibrating the feature scale to obtain a three-way normalized feature representation includes the following steps:

[0040] The reshaped single-channel 2D feature maps are input into 2D convolutional layers for feature mapping, increasing the number of channels, and then processed by batch normalization. The activation function yields the preliminary feature representation after processing.

[0041] The processed preliminary feature representation is divided into real and imaginary components in the height dimension, and a real-valued two-dimensional convolution containing real and imaginary convolution kernels is constructed. Convolution operations are performed on the real and imaginary components respectively to obtain four sets of intermediate features.

[0042] The real output component is obtained by subtracting the convolution result of the real component and the real convolution kernel and the convolution result of the imaginary component and the imaginary convolution kernel through complex value operation logic. The imaginary output component is obtained by adding the convolution result of the real component and the imaginary convolution kernel and the convolution result of the imaginary component and the real convolution kernel. The real output component and the imaginary output component are then integrated in the height dimension to generate a deep complex value feature that includes I / Q features and I / Q coupling features.

[0043] The logic for complex value operations is as follows:

[0044] ;

[0045] in They are output for the real part and output for the imaginary part, respectively. They are the real component and the imaginary component, respectively. These are the real part convolution kernel and the imaginary part convolution kernel, respectively;

[0046] Repeat the above steps three times, and input the obtained deep complex value features into two sets of pointwise convolutional units for feature recombination. Set the feature sequence length, and simultaneously map feature sequences of different lengths to the set length through two fully connected layers to obtain three normalized feature representations.

[0047] Preferably, the step of using adaptive fusion processing based on the obtained three-way feature representations, and fusing the three normalized feature representations through adaptive weight allocation to obtain the fused feature representation includes the following steps:

[0048] The three normalized feature representations are input into a two-dimensional convolutional layer for processing to obtain three-way up-dimensional features with unified dimensionality.

[0049] Based on the obtained three-way upscaling features, spatial information is aggregated through global average pooling, and the importance weights of each channel are learned through a two-layer fully connected network. At the same time, the weights are multiplied with the corresponding high-dimensional features one channel at a time for recalibration to obtain the recalibrated features. Residual connections are introduced to add the recalibrated features to the original high-dimensional features to obtain the attention-enhanced features for each channel.

[0050] The three attention enhancement features are concatenated along the channel dimension to obtain multi-channel concatenated features. These multi-channel concatenated features are then input into a pointwise convolutional layer for cross-channel information interaction and recombination, outputting the final fused feature representation.

[0051] Preferably, the step of introducing residual connections based on the recalibrated features to add the recalibrated features to the original high-dimensional features to obtain the attention-enhanced features for each path includes the following steps:

[0052] The channel importance weights learned through a two-layer fully connected network are multiplied with the corresponding high-dimensional features channel by channel to complete feature recalibration and obtain recalibrated features.

[0053] The weighting expression is shown below:

[0054] ;

[0055] in For the first Attention enhancement features for each branch, For element-wise multiplication, For the first The original high-dimensional features corresponding to each branch Assign channel importance weights;

[0056] The recalibrated features are added element-wise to the original high-dimensional features, and the summed features are output as the attention enhancement features for each path.

[0057] Preferably, the step of processing the obtained fused feature representation through the final extraction and matching recognition operation of the radio frequency fingerprint to obtain the radio frequency fingerprint recognition result includes the following steps:

[0058] The fused feature representation is input into a two-dimensional convolutional layer, and channel dimensionality reduction is performed through the two-dimensional convolutional layer to obtain the dimensionality-reduced feature representation.

[0059] The dimensionality-reduced features are input into three cascaded multi-path residual augmentation blocks. Each augmentation block contains three parallel processing paths: path one directly passes the input features to the output; path two is a local feature capture path consisting of two cascaded convolutional units; and each convolutional layer is followed by a batch normalization layer. The activation function, path three is the discriminative feature path, which is composed of depthwise separable convolution and multi-head self-attention scores concatenated. At the end of each enhancement block, the output features of the three paths are fused by adding them element by element to obtain the enhanced features.

[0060] Self-attention scores are shown below:

[0061] ;

[0062] in For the first Self-attention score of each attentional head. These are query, key, and value matrices, respectively. Represents the transpose of a matrix. This is the scaling factor;

[0063] The enhanced features are then subjected to adaptive pooling, flattening, and fully connected layer mapping in sequence, ultimately through... The function calculates the probability distribution and outputs the device identification result of the radio frequency fingerprint.

[0064] This embodiment also discloses a C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion, including: a signal and feature preprocessing module, a feature enhancement and scale alignment module, a feature adaptive fusion module, and a fingerprint recognition output module;

[0065] The signal and feature preprocessing module is used to receive baseband signals, extract effective subcarrier sequences, and extract three features based on the effective subcarrier sequences, while reshaping them into single-channel two-dimensional feature maps respectively.

[0066] The feature enhancement and scale alignment module is used to input a single-channel two-dimensional feature map into a two-dimensional convolutional layer for processing, and generate deep complex-valued features through double convolutional kernel convolution operations. At the same time, the features are mapped to a uniform length through a fully connected layer to obtain a three-way normalized feature representation.

[0067] The feature adaptive fusion module is used to increase the dimensionality of the three normalized features respectively, process them through global average pooling, and introduce residual connections to obtain attention-enhanced features. At the same time, the attention-enhanced features are concatenated and input into the pointwise convolutional layer to output the fused feature representation.

[0068] The fingerprint recognition output module is used to perform channel dimensionality reduction on the fused features, input the dimensionality-reduced features into the multi-path residual enhancement block for processing, and perform adaptive pooling and flattening operations on the enhanced features. The function outputs the device identification result corresponding to the radio frequency fingerprint.

[0069] The beneficial effects of this invention are as follows:

[0070] (1) The present invention receives the baseband signal of the physical side link broadcast channel of the transmitter through the receiver and extracts the effective subcarrier sequence. Based on the extracted effective subcarrier sequence, three channel robust features are extracted respectively. The three channel robust features are reshaped into single-channel two-dimensional feature maps. Simultaneously, complex value feature extraction and scale alignment processing are performed on the reshaped single-channel two-dimensional feature maps to extract I / Q self-features and I / Q coupling features. The feature scale is calibrated to obtain three normalized feature representations. Based on the obtained three feature representations, adaptive fusion processing is adopted, and the three normalized feature representations are fused through weight adaptive allocation to obtain fused feature representations. Finally, the obtained fused feature representations are processed through the final extraction and matching recognition operation of radio frequency fingerprint to obtain radio frequency fingerprint recognition results, which improves the accuracy and computational efficiency of device identification in complex scenarios.

[0071] (2) By utilizing three complementary features—channel equalization, complex cepstrum, and fractional domain self-ratio—this invention effectively overcomes the shortcomings of traditional single features in C-V2X high-speed mobile and multipath effect scenarios, and significantly improves the robustness of radio frequency fingerprints in complex channels. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion according to the present invention;

[0073] Figure 2 This is a schematic diagram of the overall process of the C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion described in this invention;

[0074] Figure 3 This is a structural block diagram of the complex value feature extraction and scale alignment module described in this invention;

[0075] Figure 4 This is a schematic diagram of the feature fusion module described in this invention;

[0076] Figure 5 This is a network structure diagram of the device identification module described in this invention;

[0077] Figure 6 This is a performance evaluation diagram of the cross-scene generalization capability of the present invention in a real-world scenario;

[0078] Figure 7 This is a cross-scene generalization robustness analysis diagram of the present invention in noisy real-world scenarios;

[0079] Figure 8 The present invention proposes The experimental results show the performance comparison with existing methods. Detailed Implementation

[0080] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0081] Example 1

[0082] Please see Figure 1 This embodiment discloses a C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion, including the following steps:

[0083] S1, the receiver receives the baseband signal of the physical side link broadcast channel of the transmitter and extracts the effective subcarrier sequence;

[0084] The effective subcarrier sequences include: PSSS, SSSS, and DMRS;

[0085] The process of receiving the baseband signal from the physical side link broadcast channel of the transmitter via the receiver and extracting the effective subcarrier sequence includes the following steps:

[0086] The receiver receives the baseband signal from the physical side link broadcast channel of the transmitter and obtains complete baseband signal data based on the received baseband signal.

[0087] From the complete baseband signal data, the effective subcarrier sequences corresponding to PSSS, SSSS and DMRS are separated and extracted respectively. The effective subcarrier sequences of PSSS and SSSS each contain 62 subcarriers, and the effective subcarrier sequence of DMRS contains 72 subcarriers.

[0088] S2, Based on the extracted effective subcarrier sequence, three types of channel robust features are extracted respectively, and the three types of channel robust features are reshaped into single-channel two-dimensional feature maps respectively.

[0089] The three channel robustness features include: channel equalization feature, nearest-to-nearest-symbol logarithmic difference complex cepstrum feature, and fractional-domain symbol self-ratio feature;

[0090] The process of extracting three channel-robust features based on the extracted effective subcarrier sequences, and then reshaping these three channel-robust features into single-channel two-dimensional feature maps, includes the following steps:

[0091] S21, perform least squares estimation on the effective subcarrier sequences corresponding to the extracted PSSS, SSSS and DMRS to obtain the channel equalization characteristics;

[0092] S22, the second PSSS and the first DMRS, and the first SSSS and the third DMRS in the physical side link broadcast channel subframe are taken as two sets of closest opposite symbol pairs. At the same time, 62 subcarrier sequences at the center of the first DMRS and the third DMRS are extracted. The channel influence is canceled by logarithmic difference operation and transformed to the cepstral domain by inverse Fourier transform to obtain the closest opposite symbol logarithmic difference complex cepstral features.

[0093] S23, select the effective subcarrier sequences of two PSSS and the first two DMRS in the physical side link broadcast channel, take the average and denoise them respectively, divide the averaged PSSS and DMRS into four subsequences, divide the first half subsequence with the second half subsequence to obtain the symbol self-ratio value, and map it to the fractional domain through inverse Fourier transform and fractional Fourier transform to obtain the fractional domain symbol self-ratio value characteristics.

[0094] S24. Based on the lengths of the I and Q components and the channel robustness features, the height and width of the feature map are set respectively. The extracted channel equalization features, the nearest opposite symbol logarithmic difference complex cepstral features, and the fractional domain symbol self-ratio features are reshaped into single-channel two-dimensional feature maps respectively.

[0095] The process of taking the second PSSS and the first DMRS, and the first SSSS and the third DMRS in the physical side link broadcast channel subframe as two sets of closest opposite symbol pairs, and simultaneously extracting 62 subcarrier sequences from the centers of the first DMRS and the third DMRS, canceling the channel influence through logarithmic difference operation, and transforming to the cepstral domain through inverse Fourier transform to obtain the closest opposite symbol logarithmic difference complex cepstral features includes the following steps:

[0096] In the physical side link broadcast channel subframe, the second PSSS and the first DMRS, and the first SSSS and the third DMRS are selected as two sets of closest opposite symbol pairs, and 62 subcarrier sequences at the center of the first DMRS and the third DMRS are extracted;

[0097] Logarithmic operations are performed on the sequences corresponding to the two sets of opposite symbol pairs, and then the channel influence is canceled by subtracting them pairwise. The results are then transformed into the cepstrum domain by inverse Fourier transform to obtain the most recent opposite symbol logarithmic difference complex cepstrum features.

[0098] The expression is as follows:

[0099] ;

[0100] because and Approximately equal and Since they are approximately equal, the above formula can be further expressed as:

[0101] ;

[0102] The inverse Fourier transform is shown below:

[0103] ;

[0104] in The difference between the two most recent pairs of opposite-signed logarithms. These are the channel frequency responses for the second PSSS, the first DMRS, the first SSSS, and the third DMRS, respectively. These are the sequences carrying radio frequency fingerprint information for the second PSSS, the first DMRS, the first SSSS, and the third DMRS, respectively. For the second PSSS, the effective subcarrier sequence, This is the first DMRS subcarrier sequence. This is the first valid subcarrier sequence of SSSS. This is the third DMRS subcarrier sequence. For two most recent log-difference complex cepstral differences with different signs, This is the inverse Fourier transform.

[0105] The steps involved in selecting two effective subcarrier sequences of the physical side link broadcast channel (PSSS) and the first two DMRS, averaging and denoising them, dividing each of the averaged PSSS and DMRS into four subsequences, dividing the first half of the subsequence by the second half to obtain the symbol self-ratio, and mapping it to the fractional domain through inverse Fourier transform and fractional Fourier transform to obtain the fractional domain symbol self-ratio characteristics include the following steps:

[0106] Select the effective subcarrier sequences of the two PSSS and the first two DMRS in the physical side link broadcast channel, take the average noise reduction, divide them into four subsequences, and divide the first half subsequence with the second half subsequence to obtain the symbol self-ratio value.

[0107] Perform inverse Fourier transform and fractional Fourier transform on the sign self-ratio in sequence, and map it to the fractional domain to obtain the fractional domain sign self-ratio characteristics;

[0108] The expression for the symbolic self-ratio is shown below:

[0109] ;

[0110] The characteristics of the sign-based ratio in the fractional domain are shown below:

[0111] ;

[0112] in The sign-to-value ratios of PSSS and DMRS are given. It is a four-segment subsequence that is equally divided. The fractional sign ratios of PSSS and DMRS are respectively. for Fractional Fourier transform, This is the transformation kernel function.

[0113] S3, perform complex value feature extraction and scale alignment processing on the reshaped single-channel two-dimensional feature map, extract the I / Q features and I / Q coupling features, and complete the feature scale calibration to obtain the three-way normalized feature representation;

[0114] The process of performing complex value feature extraction and scale alignment on the reconstructed single-channel two-dimensional feature map, extracting I / Q features and I / Q coupling features, and calibrating the feature scale to obtain a three-way normalized feature representation includes the following steps:

[0115] The reshaped single-channel 2D feature maps are input into 2D convolutional layers for feature mapping, increasing the number of channels, and then processed by batch normalization. The activation function yields the preliminary feature representation after processing.

[0116] The processed preliminary feature representation is divided into real and imaginary components in the height dimension, and a real-valued two-dimensional convolution containing real and imaginary convolution kernels is constructed. Convolution operations are performed on the real and imaginary components respectively to obtain four sets of intermediate features.

[0117] The real output component is obtained by subtracting the convolution result of the real component and the real convolution kernel and the convolution result of the imaginary component and the imaginary convolution kernel through complex value operation logic. The imaginary output component is obtained by adding the convolution result of the real component and the imaginary convolution kernel and the convolution result of the imaginary component and the real convolution kernel. The real output component and the imaginary output component are then integrated in the height dimension to generate a deep complex value feature that includes I / Q features and I / Q coupling features.

[0118] The logic for complex value operations is as follows:

[0119] ;

[0120] in They are output for the real part and output for the imaginary part, respectively. They are the real component and the imaginary component, respectively. These are the real part convolution kernel and the imaginary part convolution kernel, respectively;

[0121] Repeat the above steps three times, and input the obtained deep complex value features into two sets of pointwise convolutional units for feature recombination. Set the feature sequence length, and simultaneously map feature sequences of different lengths to the set length through two fully connected layers to obtain three normalized feature representations.

[0122] S4. Based on the obtained three-way feature representations, an adaptive fusion process is adopted, and the three normalized feature representations are fused through adaptive weight allocation to obtain a fused feature representation.

[0123] The process of obtaining the three-way feature representation by adaptive fusion processing and fusing the three normalized feature representations through adaptive weight allocation to obtain the fused feature representation includes the following steps:

[0124] The three normalized feature representations are input into a two-dimensional convolutional layer for processing to obtain three-way up-dimensional features with unified dimensionality.

[0125] Based on the obtained three-way upscaling features, spatial information is aggregated through global average pooling, and the importance weights of each channel are learned through a two-layer fully connected network. At the same time, the weights are multiplied with the corresponding high-dimensional features one channel at a time for recalibration to obtain the recalibrated features. Residual connections are introduced to add the recalibrated features to the original high-dimensional features to obtain the attention-enhanced features for each channel.

[0126] The global information aggregation is shown below:

[0127] ;

[0128] in It is the first The first input branch Preliminary characteristics of the channel, It is the first The first input branch Global characteristics of the channel;

[0129] The three attention enhancement features are concatenated along the channel dimension to obtain multi-channel concatenated features. These multi-channel concatenated features are then input into a pointwise convolutional layer for cross-channel information interaction and recombination, outputting the final fused feature representation.

[0130] The process of introducing residual connections based on the recalibrated features to add the recalibrated features to the original high-dimensional features, thereby obtaining the attention-enhanced features for each path, includes the following steps:

[0131] The channel importance weights learned through a two-layer fully connected network are multiplied with the corresponding high-dimensional features channel by channel to complete feature recalibration and obtain recalibrated features.

[0132] The weighting expression is shown below:

[0133] ;

[0134] in For the first Attention enhancement features for each branch, For element-wise multiplication, Assign channel importance weights;

[0135] The recalibrated features are added element-wise to the original high-dimensional features, and the summed features are output as the attention enhancement features for each path.

[0136] S5 processes the obtained fused feature representation through the final extraction and matching recognition operation of the radio frequency fingerprint to obtain the radio frequency fingerprint recognition result.

[0137] The process of processing the obtained fused feature representation through the final extraction and matching recognition operation of the radio frequency fingerprint to obtain the radio frequency fingerprint recognition result includes the following steps:

[0138] The fused feature representation is input into a two-dimensional convolutional layer, and channel dimensionality reduction is performed through the two-dimensional convolutional layer to obtain the dimensionality-reduced feature representation.

[0139] The dimensionality-reduced features are input into three cascaded multi-path residual augmentation blocks. Each augmentation block contains three parallel processing paths: path one directly passes the input features to the output; path two is a local feature capture path consisting of two cascaded convolutional units; and each convolutional layer is followed by a batch normalization layer. The activation function, path three is the discriminative feature path, which is composed of depthwise separable convolution and multi-head self-attention scores concatenated. At the end of each enhancement block, the output features of the three paths are fused by adding them element by element to obtain the enhanced features.

[0140] Self-attention scores are shown below:

[0141] ;

[0142] in For the first Self-attention score of each attentional head. These are query, key, and value matrices, respectively. Represents the transpose of a matrix. This is the scaling factor;

[0143] The enhanced features are then subjected to adaptive pooling, flattening, and fully connected layer mapping in sequence, ultimately through... The function calculates the probability distribution and outputs the device identification result of the radio frequency fingerprint.

[0144] Example 2

[0145] This embodiment also discloses a system for a C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion, including: a signal and feature preprocessing module, a feature enhancement and scale alignment module, a feature adaptive fusion module, and a fingerprint recognition output module;

[0146] The signal and feature preprocessing module is used to receive baseband signals, extract effective subcarrier sequences, and extract three features based on the effective subcarrier sequences, while reshaping them into single-channel two-dimensional feature maps respectively.

[0147] The feature enhancement and scale alignment module is used to input a single-channel two-dimensional feature map into a two-dimensional convolutional layer for processing, and generate deep complex-valued features through double convolutional kernel convolution operations. At the same time, the features are mapped to a uniform length through a fully connected layer to obtain a three-way normalized feature representation.

[0148] The feature adaptive fusion module is used to increase the dimensionality of the three normalized features respectively, process them through global average pooling, and introduce residual connections to obtain attention-enhanced features. At the same time, the attention-enhanced features are concatenated and input into the pointwise convolutional layer to output the fused feature representation.

[0149] The fingerprint recognition output module is used to perform channel dimensionality reduction on the fused features, input the dimensionality-reduced features into the multi-path residual enhancement block for processing, and perform adaptive pooling and flattening operations on the enhanced features. The function outputs the device identification result corresponding to the radio frequency fingerprint.

[0150] To verify the effectiveness and robustness of the method proposed in this invention, performance comparison experiments and analyses were conducted. (See attached figure.) Figure 6 To be continued Figure 8 The experimental results of the method of the present invention and the comparative object are described in detail below.

[0151] exist Figure 6 , Figure 7 and Figure 8 The specific meanings of the comparison objects involved in the experimental results are as follows:

[0152] The "S.2" refers to a stationary scene at a line of sight.

[0153] The "S.3" refers to: line-of-sight movement scenario;

[0154] The “S.4” refers to: a non-line-of-sight static scene;

[0155] The "S.5" refers to: non-line-of-sight movement scenario;

[0156] The "S.6" refers to: outdoor mobile scenario;

[0157] The "feature 1" is: channel equalization feature;

[0158] The "feature 2" is: the most recent log-difference complex cepstral feature of the opposite sign;

[0159] The "feature 3" is: fractional domain sign self-ratio feature;

[0160] The "Comparison Method 1" is: the advanced feature fusion model McAFF;

[0161] The "comparison method 2" refers to the MFF model, which is fused by direct splicing.

[0162] The “comparison method 3” is: a channel robust feature extraction method based on spectral ratio;

[0163] The “Comparison Method 4” is a channel robust feature extraction method based on time correlation.

[0164] It should be noted that the above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion, characterized in that, Includes the following steps: S1, the receiver receives the baseband signal of the physical side link broadcast channel of the transmitter and extracts the effective subcarrier sequence; The effective subcarrier sequences include: PSSS, SSSS, and DMRS; S2, based on the extracted effective subcarrier sequences, three types of features with channel robustness are extracted respectively, and the three types of extracted features with channel robustness are reconstructed into single-channel two-dimensional feature maps, including the following steps: S21, perform least squares estimation on the effective subcarrier sequences corresponding to the extracted PSSS, SSSS and DMRS to obtain the channel equalization characteristics; S22, the second PSSS and the first DMRS, and the first SSSS and the third DMRS in the physical side link broadcast channel subframe are taken as two sets of closest opposite symbol pairs. At the same time, 62 subcarrier sequences at the center of the first DMRS and the third DMRS are extracted. The channel influence is canceled by logarithmic difference operation and transformed to the cepstral domain by inverse Fourier transform to obtain the closest opposite symbol logarithmic difference complex cepstral features. S23, select the effective subcarrier sequences of two PSSS and the first two DMRS in the physical side link broadcast channel, take the average and denoise them respectively, divide the averaged PSSS and DMRS into four subsequences, divide the first half subsequence with the second half subsequence to obtain the symbol self-ratio value, and map it to the fractional domain through inverse Fourier transform and fractional Fourier transform to obtain the fractional domain symbol self-ratio value characteristics. S24. Based on the lengths of the I and Q components and the channel robustness features, the height and width of the feature map are set respectively. The extracted channel equalization features, the nearest opposite symbol log difference complex cepstral features and the fractional domain symbol self-ratio features are reshaped into single-channel two-dimensional feature maps respectively. The three channel robustness features include: channel equalization feature, nearest-to-nearest-symbol logarithmic difference complex cepstrum feature, and fractional-domain symbol self-ratio feature; S3, perform complex value feature extraction and scale alignment processing on the reshaped single-channel two-dimensional feature map, extract the I / Q features and I / Q coupling features, and complete the feature scale calibration to obtain the three-way normalized feature representation; S4. Based on the obtained three-way feature representations, an adaptive fusion process is adopted, and the three normalized feature representations are fused through adaptive weight allocation to obtain a fused feature representation. S5 processes the obtained fused feature representation through the final extraction and matching recognition operation of the radio frequency fingerprint to obtain the radio frequency fingerprint recognition result.

2. The C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion according to claim 1, characterized in that, The process of receiving the baseband signal from the physical side link broadcast channel of the transmitter via the receiver and extracting the effective subcarrier sequence includes the following steps: The receiver receives the baseband signal from the physical side link broadcast channel of the transmitter and obtains complete baseband signal data based on the received baseband signal. From the complete baseband signal data, the effective subcarrier sequences corresponding to PSSS, SSSS and DMRS are separated and extracted respectively. The effective subcarrier sequences of PSSS and SSSS each contain 62 subcarriers, and the effective subcarrier sequence of DMRS contains 72 subcarriers.

3. The C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion according to claim 1, characterized in that, The process of taking the second PSSS and the first DMRS, and the first SSSS and the third DMRS in the physical side link broadcast channel subframe as two sets of closest opposite symbol pairs, and simultaneously extracting 62 subcarrier sequences from the centers of the first DMRS and the third DMRS, canceling the channel influence through logarithmic difference operation, and transforming to the cepstral domain through inverse Fourier transform to obtain the closest opposite symbol logarithmic difference complex cepstral features includes the following steps: In the physical side link broadcast channel subframe, the second PSSS and the first DMRS, and the first SSSS and the third DMRS are selected as two sets of closest opposite symbol pairs, and 62 subcarrier sequences at the center of the first DMRS and the third DMRS are extracted; Logarithmic operations are performed on the sequences corresponding to the two sets of opposite symbol pairs, and then the channel influence is canceled out by pairwise subtraction. The results are then transformed into the cepstrum domain using an inverse Fourier transform to obtain the most recent opposite symbol logarithmic difference complex cepstrum features. The expression is as follows: ； because and Approximately equal and Since they are approximately equal, the above formula can be further expressed as: ； The inverse Fourier transform is shown below: ； in The difference between the two most recent pairs of opposite-signed logarithms. These are the channel frequency responses for the second PSSS, the first DMRS, the first SSSS, and the third DMRS, respectively. These are the sequences carrying radio frequency fingerprint information for the second PSSS, the first DMRS, the first SSSS, and the third DMRS, respectively. For the second PSSS, the effective subcarrier sequence, This is the first DMRS subcarrier sequence. This is the first valid subcarrier sequence of SSSS. This is the third DMRS subcarrier sequence. For two most recent log-difference complex cepstral differences with different signs, This is the inverse Fourier transform.

4. The C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion according to claim 1, characterized in that, The steps involved in selecting two effective subcarrier sequences of the physical side link broadcast channel (PSSS) and the first two DMRS, averaging and denoising them, dividing each of the averaged PSSS and DMRS into four subsequences, dividing the first half of the subsequence by the second half to obtain the symbol self-ratio, and mapping it to the fractional domain through inverse Fourier transform and fractional Fourier transform to obtain the fractional domain symbol self-ratio characteristics include the following steps: Select the effective subcarrier sequences of the two PSSS and the first two DMRS in the physical side link broadcast channel, take the average noise reduction, divide them into four subsequences, and divide the first half subsequence with the second half subsequence to obtain the symbol self-ratio value. Perform inverse Fourier transform and fractional Fourier transform on the sign self-ratio in sequence, and map it to the fractional domain to obtain the fractional domain sign self-ratio characteristics; The expression for the symbolic self-ratio is shown below: ； The characteristics of the sign-based ratio in the fractional domain are as follows: ； in The sign-based self-ratio values ​​for PSSS and DMRS are given. It is a four-segment subsequence that is equally divided. The fractional sign ratios of PSSS and DMRS are respectively. for Fractional Fourier transform, This is the transformation kernel function.

5. The C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion according to claim 1, characterized in that, The process of performing complex value feature extraction and scale alignment on the reconstructed single-channel two-dimensional feature map, extracting I / Q features and I / Q coupling features, and calibrating the feature scale to obtain a three-way normalized feature representation includes the following steps: The reshaped single-channel 2D feature maps are input into 2D convolutional layers for feature mapping, increasing the number of channels, and then processed by batch normalization. The activation function yields the preliminary feature representation after processing. The processed preliminary feature representation is divided into real and imaginary components in the height dimension, and a real-valued two-dimensional convolution containing real and imaginary convolution kernels is constructed. Convolution operations are performed on the real and imaginary components respectively to obtain four sets of intermediate features. The real output component is obtained by subtracting the convolution result of the real component and the real convolution kernel and the convolution result of the imaginary component and the imaginary convolution kernel through complex value operation logic. The imaginary output component is obtained by adding the convolution result of the real component and the imaginary convolution kernel and the convolution result of the imaginary component and the real convolution kernel. The real output component and the imaginary output component are then integrated in the height dimension to generate a deep complex value feature that includes I / Q features and I / Q coupling features. Repeat the above steps three times, and input the obtained deep complex value features into two sets of pointwise convolutional units for feature recombination. Set the feature sequence length, and simultaneously map feature sequences of different lengths to the set length through two fully connected layers to obtain three normalized feature representations.

6. The C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion according to claim 1, characterized in that, The process of obtaining the three-way feature representation by adaptive fusion processing and fusing the three normalized feature representations through adaptive weight allocation to obtain the fused feature representation includes the following steps: The three normalized feature representations are input into a two-dimensional convolutional layer for processing to obtain three-way up-dimensional features with unified dimensionality. Based on the obtained three-way up-dimensional features, spatial information is aggregated through global average pooling, and the importance weights of each channel are learned through a two-layer fully connected network. At the same time, the weights are multiplied with the corresponding high-dimensional features channel by channel for recalibration to obtain the recalibrated features. Based on the recalibrated features, residual connections are introduced to add the recalibrated features to the original high-dimensional features, resulting in attention-enhanced features for each path. The three attention enhancement features are concatenated along the channel dimension to obtain multi-channel concatenated features. These multi-channel concatenated features are then input into a pointwise convolutional layer for cross-channel information interaction and recombination, outputting the final fused feature representation.

7. The C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion according to claim 1, characterized in that, The process of processing the obtained fused feature representation through the final extraction and matching recognition operation of the radio frequency fingerprint to obtain the radio frequency fingerprint recognition result includes the following steps: The fused feature representation is input into a two-dimensional convolutional layer, and channel dimensionality reduction is performed through the two-dimensional convolutional layer to obtain the dimensionality-reduced feature representation; The dimensionality-reduced features are input into three cascaded multi-path residual augmentation blocks. Each augmentation block contains three parallel processing paths: path one directly passes the input features to the output; path two is a local feature capture path consisting of two cascaded convolutional units; and each convolutional layer is followed by a batch normalization layer. The activation function, path three is the discriminative feature path, which is composed of depthwise separable convolution and multi-head self-attention scores concatenated. At the end of each enhancement block, the output features of the three paths are fused by adding them element by element to obtain the enhanced features. The enhanced features are then subjected to adaptive pooling, flattening, and fully connected layer mapping in sequence, ultimately through... The function calculates the probability distribution and outputs the device identification result of the radio frequency fingerprint.

8. The C-V2X radio frequency fingerprint recognition method based on multi-channel robust feature fusion according to claim 1, characterized in that, The method includes: a signal and feature preprocessing module, a feature enhancement and scale alignment module, a feature adaptive fusion module, and a fingerprint recognition output module; The signal and feature preprocessing module is used to receive baseband signals, extract effective subcarrier sequences, and extract three features based on the effective subcarrier sequences, while reshaping them into single-channel two-dimensional feature maps respectively. The feature enhancement and scale alignment module is used to input a single-channel two-dimensional feature map into a two-dimensional convolutional layer for processing, and generate deep complex-valued features through double convolutional kernel convolution operations. At the same time, the features are mapped to a uniform length through a fully connected layer to obtain a three-way normalized feature representation. The feature adaptive fusion module is used to increase the dimensionality of the three normalized features respectively, process them through global average pooling, and introduce residual connections to obtain attention-enhanced features. At the same time, the attention-enhanced features are concatenated and input into the pointwise convolutional layer to output the fused feature representation. The fingerprint recognition output module is used to perform channel dimensionality reduction on the fused features, input the dimensionality-reduced features into the multi-path residual enhancement block for processing, and perform adaptive pooling and flattening operations on the enhanced features. The function outputs the device identification result corresponding to the radio frequency fingerprint.

Images