A method for radio frequency fingerprinting of broadcast frames in vehicle-to-everything (V2X) networks based on LMMSE channel estimation

Abstract

This invention provides a method for radio frequency fingerprinting of vehicular network broadcast frames based on LMMSE channel estimation, comprising: obtaining subcarrier data in the resource grid based on the acquired signal of the physical side link broadcast channel; performing root mean square delay spreading and channel autocorrelation matrix construction based on the subcarrier data, and combining prior signal-to-noise ratio regularization and time-domain windowing operation to obtain channel estimation values; obtaining initial radio frequency fingerprint features through channel equalization and channel estimation values, and using an improved neural network and initial radio frequency fingerprint features to classify and identify vehicular network devices. This method effectively considers the impact of noise and channel on fingerprints, achieving the goal of more accurately extracting fingerprints of different devices in complex environments.

Description

Technical Field

[0001] This invention belongs to the field of V2X communication technology, and in particular relates to a method for radio frequency fingerprinting of vehicle-to-everything (V2X) broadcast frames based on LMMSE channel estimation. Background Technology

[0002] With the rapid development of Intelligent Transportation Systems (ITS), Cellular Vehicle-to-Everything (C-V2X) technology has become one of the key technologies driving intelligent transportation, and has been widely used in areas such as vehicle cooperative perception, emergency warning, and traffic efficiency optimization. However, in the C-V2X communication environment, problems such as unauthorized access of terminal devices, malicious behavior, and device identity forgery are becoming increasingly prominent, seriously threatening network security and reliability. Physical layer security offers a new approach through radio frequency fingerprinting technology. Radio frequency fingerprinting utilizes inherent hardware defects (such as I / Q imbalance and power amplifier nonlinearity) as unclonable physical identifiers, circumventing the problem of computational complexity dependence. Furthermore, its unforgeability and uniqueness make it an excellent medium for device identification and authentication.

[0003] In complex wireless channel environments, the robustness of radio frequency (RF)-based systems is compromised. Frequency-selective and time-varying channels struggle to extract clean and stable RFF features decoupled from channel characteristics, leading to poor performance in multipath and mobile scenarios. Channel interference can cause fingerprint classification accuracy to drop by up to 80%. Although Channel State Information (CSI) itself can serve as a potential medium for device authentication, environmental semantic-based physical layer authentication mechanisms face performance degradation in dynamic scenarios (such as transmitter movement or the proximity of malicious and legitimate devices). To address channel sensitivity, existing research focuses on designing channel-robust RF fingerprint feature extraction methods. Mainstream approaches focus on extracting inherent transmitter features independent of the channel, such as using Carrier Frequency Offset (CFO) as a device identification feature. However, CFO suffers from insufficient temporal stability and low class discrimination, making it difficult to achieve reliable device classification performance in multipath fading and multi-device coexistence scenarios. With the widespread application of deep learning technology in RF fingerprint recognition, data-driven methods based on adversarial training have become a research hotspot. For example, adversarial examples can be used as augmented data to constrain feature extractors through strong regularization, but this would sacrifice model sensitivity and standard accuracy.

[0004] However, the aforementioned methods are mostly designed for narrowband or low-mobility scenarios, making them difficult to adapt to the dynamic requirements of cellular vehicle-to-everything (V2X) networks. Some attempts have tried to mitigate channel interference using data augmentation and artificial noise addition methods, but the former significantly increases model training overhead, while the latter, based on the linear superposition assumption of "received signal = transmitted signal + RFF," suffers from model mismatch with the nonlinear coupling characteristics of channel distortion and hardware impairment in physical layer signal transmission. Furthermore, it is worth noting that while systems like Wi-Fi and LoRa can achieve channel compensation through protocol-specific frame structure features, their methodologies are limited by the heterogeneity of wireless communication standards and cannot be directly transferred to the C-V2X system architecture. In addition, the differential constellation trajectory map construction method based on the transient characteristics of the physical random access channel, although extracting RFF features during the transient on / off phases using a multi-channel convolutional neural network, shows a significant decrease in classification accuracy for persistent symbol features such as the demodulation reference signal (DMRS). Achieving dynamic channel robustness while maintaining the inherent characteristics of RFF is crucial for the reliability of C-V2X security authentication. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a radio frequency fingerprint recognition method for vehicle-to-everything (V2X) broadcast frames based on LMMSE channel estimation, which at least partially solves the problems of incomplete fingerprint extraction and low classification accuracy in C-V2X systems in the prior art.

[0006] This disclosure provides a method for radio frequency fingerprinting of vehicle-to-everything (V2X) broadcast frames based on LMMSE channel estimation, including:

[0007] Subcarrier data in the resource grid is obtained based on the signal from the acquired physical sidelink broadcast channel;

[0008] Based on subcarrier data, root mean square delay spread and channel autocorrelation matrix construction are performed. Combined with prior signal-to-noise ratio regularization and time-domain windowing operation, the channel estimate is obtained.

[0009] Initial radio frequency fingerprint features are obtained through channel equalization and channel estimation, and the improved neural network and initial radio frequency fingerprint features are used to classify and identify vehicle networking devices.

[0010] Optionally, subcarrier data in the resource grid is obtained based on the signal of the acquired physical side link broadcast channel, including: signal detection, frame synchronization, carrier frequency offset, and resource grid demodulation;

[0011] The signal detection method uses block energy gradient and threshold determination to locate the signal starting point by calculating the energy ratio of adjacent data blocks.

[0012] The frame synchronization utilizes the periodic characteristics of the master synchronization signal for coarse synchronization and searches for signal peak points through amplitude normalization cross-correlation calculation.

[0013] The carrier frequency offset compensation utilizes the characteristic that two consecutive primary synchronization symbols are the same and two consecutive secondary synchronization symbols are the same, and obtains the frequency offset estimate by using the signal autocorrelation of the physical side link broadcast channel.

[0014] The resource grid demodulation utilizes the structural characteristics of the cyclic prefix and eliminates the effects of time domain offset and subcarrier misalignment through phase and half-subcarrier compensation to obtain frequency domain resources.

[0015] Optionally, the signal detection includes:

[0016] Define the signal block length as Traverse the signals of the physical side link broadcast channel, according to the block length Divide the data into non-overlapping segments and calculate the energy block by block. :

[0017] ,

[0018] in For the first Grid data blocks, For sampling point index, for transpose, It is a natural number.

[0019] Optionally, the carrier frequency offset compensation includes:

[0020] Extract the signal segment, calculate the conjugate cross-correlation phase difference, and estimate the frequency offset:

[0021]

[0022] in, For frequency offset, , The number of points in the Fourier transform. and The relevant segment of the main synchronization signal and The relevant segment for the auxiliary synchronization symbol, where n is a natural number.

[0023] Optionally, the resource grid demodulation includes:

[0024] Determine the starting position of the Fourier transform, and then extract the length as... The time-domain sample segment, with a length of Time-domain sample segments and factors Multiply, then perform a Fourier transform on the resulting time-domain signal to convert the time-domain sample segment to the frequency domain, factoring... Used for half-subcarrier frequency offset compensation.

[0025] Optionally, the step of performing root mean square delay spreading and channel autocorrelation matrix construction based on subcarrier data, combined with prior signal-to-noise ratio regularization and time-domain windowing operations, to obtain the channel estimate includes:

[0026] Estimate channel power to characterize signal energy;

[0027] The channel autocorrelation matrix is ​​constructed based on the estimated channel power and root mean square delay spread.

[0028] Optionally, the channel autocorrelation matrix is ​​constructed based on the estimated channel power and the delay-related phase offset, which is obtained based on the frequency difference between subcarriers.

[0029] Optionally, by combining prior signal-to-noise ratio regularization and time-domain windowing, channel estimates are obtained, including:

[0030] Generate dynamically adjusted regularization terms Regularization term This is used to reduce the impact of regularization when the prior signal-to-noise ratio is high, and to increase the regularization strength when the prior signal-to-noise ratio is low.

[0031] Optionally, the step of performing root mean square delay spreading and channel autocorrelation matrix construction based on subcarrier data, combined with prior signal-to-noise ratio regularization and time-domain windowing operations, to obtain the channel estimate includes:

[0032] In low signal-to-noise ratio scenarios, the default setting is the root mean square delay spread value;

[0033] The initial frequency domain channel response is averaged using multiple symbols to smooth out the noise effects.

[0034] The data differences in the channel autocorrelation matrix are eliminated by symmetry processing, and the main path is preserved by windowing operation to suppress noise and radio frequency fingerprint interference.

[0035] Optionally, initial radio frequency fingerprint features are obtained through channel equalization and channel estimation, and the improved neural network and initial radio frequency fingerprint features are used to classify and identify vehicular network devices. The improved neural network is an improved ShuffleNet V2 network, which includes:

[0036] A convolutional block attention module is integrated after the final convolutional layer, and key spectral features are dynamically focused through joint optimization of channel attention and spatial attention.

[0037] The channel attention module extracts channel features through global average pooling and global max pooling, and generates channel weights by combining them with a fully connected layer. The spatial attention module generates spatial features on the feature map after channel attention weighting by average pooling and max pooling along the channel dimension, and then generates spatial weights by using one-dimensional convolution.

[0038] This invention provides a radio frequency fingerprinting method for vehicle-to-everything (V2X) broadcast frames based on LMMSE channel estimation. It processes the frequency domain signal to obtain a channel estimate, uses channel equalization and the channel estimate to obtain initial radio frequency fingerprint features, and then utilizes a neural network for device classification and identification. This method considers the correlation between time and frequency, maintains the robustness of LMMSE, reduces excessive reliance on pilot data in high-noise environments, and suppresses noise amplification effects. Addressing the resource constraints of C-V2X, based on an analysis of the model complexity in this field, a lightweight model is constructed, exhibiting robustness to different scenarios and noise levels, and possessing the potential for on-site training and adjustment. It effectively considers the impact of noise and channel on fingerprints, thereby achieving more accurate extraction of fingerprints for different devices in complex environments. Attached Figure Description

[0039] The above and other objects, features and advantages of this disclosure will become more apparent from the accompanying drawings, in which like reference numerals generally denote like parts.

[0040] Figure 1 A flowchart of a vehicle-to-everything (V2X) broadcast frame radio frequency fingerprinting method based on LMMSE channel estimation provided in this embodiment of the disclosure;

[0041] Figure 2 This is a schematic diagram of the physical side link broadcast frame structure provided in an embodiment of the present disclosure;

[0042] Figure 3 This is a schematic diagram illustrating the cross-correlation of received signals with a standard PSSS sequence pair provided in an embodiment of this disclosure;

[0043] Figure 4a This is a schematic diagram of an unprocessed radio frequency fingerprint provided in an embodiment of this disclosure;

[0044] Figure 4b This is a schematic diagram of channel processing radio frequency fingerprinting provided in an embodiment of this disclosure;

[0045] Figure 5 An overall architecture diagram of the improved ShuffleNet V2 module provided in this disclosure embodiment;

[0046] Figure 6 The inverse residual basic module of the neural network provided in the embodiments of this disclosure;

[0047] Figure 7 The inverse residual downsampling module of the neural network provided in the embodiments of this disclosure;

[0048] Figure 8 This is a comparison chart of accuracy under different signal-to-noise ratios provided in the embodiments of this disclosure. Detailed Implementation

[0049] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0050] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0051] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0052] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The illustrations only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0053] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0054] This embodiment provides a method for radio frequency fingerprinting of vehicular network broadcast frames based on LMMSE channel estimation. Specifically, it aims to optimize channel response estimation by combining the channel covariance matrix and noise statistics. Initial fingerprint features are generated by removing the channel response, and the ShuffleNet V2 network architecture is improved to optimize it for low signal-to-noise ratio and dynamic channel conditions, while maintaining high feature extraction capabilities. During training, training data collected from different scenarios are integrated, utilizing diverse samples to enhance model generalization ability. Furthermore, by increasing the information richness of features, the computational load and parameter count are effectively reduced, thereby improving computational efficiency while maintaining detection accuracy and achieving network lightweighting. It can more accurately classify devices in low signal-to-noise ratio stationary and moving scenarios.

[0055] The received signal in this embodiment is the signal of the physical side link broadcast channel.

[0056] For ease of understanding, this embodiment discloses a radio frequency fingerprinting method for vehicular network broadcast frames based on LMMSE channel estimation.

[0057] This includes obtaining subcarrier data in the resource grid based on the signals from the acquired physical sidelink broadcast channels;

[0058] Based on subcarrier data, root mean square delay spread and channel autocorrelation matrix construction are performed. Combined with prior signal-to-noise ratio regularization and time-domain windowing operation, the channel estimate is obtained.

[0059] The process of performing root mean square delay spreading and channel autocorrelation matrix construction based on subcarrier data, combined with prior signal-to-noise ratio regularization and time-domain windowing, yields channel estimation, including:

[0060] The estimated root mean square delay spread is used to construct the channel autocorrelation matrix;

[0061] A prior signal-to-noise ratio regularization term is introduced to dynamically adjust the regularization strength in LMMSE estimation;

[0062] The frequency domain channel response is transformed to the time domain by inverse Fourier transform, a windowing operation is performed to suppress noise and non-channel components, and then the frequency domain is restored by Fourier transform to obtain the final channel estimate.

[0063] Initial radio frequency fingerprint features are obtained through channel equalization and channel estimation. An improved neural network and these initial radio frequency fingerprint features are then used to classify and identify connected vehicle devices.

[0064] This allows for the removal of channel effects using channel equalization techniques, extraction of initial radio frequency fingerprint features, and the channel effects to be obtained based on the final channel estimation.

[0065] The subcarrier data in the resource grid is obtained based on the signal of the acquired physical side link broadcast channel, including: signal detection, frame synchronization, carrier frequency offset and resource grid demodulation;

[0066] The signal detection method uses block energy gradient and threshold determination to locate the signal starting point by calculating the energy ratio of adjacent data blocks.

[0067] The frame synchronization utilizes the periodic characteristics of the master synchronization signal for coarse synchronization and searches for signal peak points through amplitude normalization cross-correlation calculation.

[0068] The carrier frequency offset compensation utilizes the characteristic that two consecutive primary synchronization symbols are the same and two consecutive secondary synchronization symbols are the same, and obtains the frequency offset estimate by using the signal autocorrelation of the physical side link broadcast channel.

[0069] The resource grid demodulation utilizes the structural characteristics of the cyclic prefix and eliminates the effects of time domain offset and subcarrier misalignment through phase and half-subcarrier compensation to obtain frequency domain resources.

[0070] Signal detection uses a dynamic interception strategy to determine the effective signal segment, ensuring the accuracy of the signal start point.

[0071] Frame synchronization is based on the strong periodicity of PSSS and SSSS signals, combined with fine synchronization of CP, to verify the standard subframe length of 2048 and optimize synchronization accuracy.

[0072] Carrier frequency offset compensation utilizes the same characteristics of synchronization symbols and combines the phase difference between the CP and the data tail to achieve high-precision frequency offset estimation and compensation.

[0073] like Figure 1 As shown, the process includes: Step S1: The receiver oversamples and downconverts the signal to obtain the time-domain baseband received signal. Preprocessing is performed on this signal, and after time synchronization, carrier frequency offset estimation, and compensation, a time-domain received signal carrying RF fingerprint and channel information is obtained. The signal is then demodulated using a resource grid to obtain the frequency-domain resource. Step S2: A high-precision frequency-domain channel response is generated using Root Mean Square Delay Spread (RMS Delay Spread) and a channel autocorrelation matrix, combined with prior signal-to-noise ratio regularization and time-domain windowing, effectively suppressing noise and multipath interference. Step S3: Initial RF fingerprint features are extracted through channel equalization, and an improved lightweight neural network is used for device classification and identification. This embodiment exhibits high robustness in low signal-to-noise ratio and dynamic scenarios, making it suitable for device authentication in vehicle-to-everything (C-V2X) scenarios.

[0074] The LMMSE channel estimate is constructed using the root mean square delay spread and the channel autocorrelation matrix, including:

[0075] In low signal-to-noise ratio scenarios, the default RMS delay spread value is set to ensure estimation stability;

[0076] Multi-symbol averaging is performed on the initial frequency domain channel response to smooth out noise effects;

[0077] By eliminating data differences in the channel autocorrelation matrix through symmetry processing and preserving the main path through windowing operations, noise and RFF interference are suppressed, thus obtaining accurate channel estimates.

[0078] In the vehicular network broadcast frame radio frequency fingerprinting method based on LMMSE channel estimation in this embodiment, as follows: Figure 2 and Figure 3 As shown, the Primary Sidelink Synchronization Signal (PSSS) and the Secondary Sidelink Synchronization Signal (SSSS) are configured in symbols 2-3 and 12-13 respectively to achieve fast time-frequency synchronization; the DMRS cluster (symbols 5, 7, and 10) provides high-resolution channel state information (CSI); the GP is deployed in symbol 14 to absorb multipath energy; the remaining symbols are PSBCH symbols. The signal from the Physical Sidelink Broadcast Channel (PSBCH) is received, and subcarrier data in the resource grid is obtained through signal detection, frame synchronization, carrier frequency offset (CFO), and resource grid demodulation, specifically including:

[0079] Signal detection employs a method based on block energy gradient and threshold determination to locate the signal start point, and uses a dynamic truncation strategy to determine the effective signal segment, specifically:

[0080] First, define the signal block length as... traverse and receive signals ,in, This refers to the baseband signal length. Based on block length... Divide the data into non-overlapping segments and calculate the energy block by block. :

[0081] ,

[0082] in For the first For a given data block, if the energy ratio between adjacent blocks exceeds a threshold T, the signal's starting point can be determined to be located at... place, For sampling point index, for The transpose of the signal can be used to locate and detect the target signal segment;

[0083] Subsequently, for frame synchronization, due to the special characteristics of PSBCH, the number of Fast Fourier Transform (FFT) points is set in the initial stage. And define the cyclic prefix length vector. The CP lengths for symbols 1 and 8 are 160 samples, while the rest are 144 samples, consistent with the normal CP configuration of a sidechain. Coarse synchronization is achieved using the periodic characteristics of the master synchronization signal (PSSS) by loading a locally generated normalized PSSS sequence. The peak point is searched by performing amplitude-normalized cross-correlation on the received signal, using the following formula:

[0084] ,

[0085] in For the relevant window length, For sampling point index, The correlation coefficient is calculated by searching for the peak point of the correlation curve. and The interval was verified to meet the standard subframe length of 2048. Subsequently, the synchronization accuracy was further optimized based on fine synchronization using the cyclic prefix (CP). A local search range was defined near the coarse synchronization point, and the CP cross-correlation of M consecutive OFDM symbols was accumulated. The offset with the largest cumulative value is selected as the final synchronization point. :

[0086] ,

[0087] For the first The starting point of an OFDM symbol, The total number of symbols.

[0088] Furthermore, carrier frequency offset compensation utilizes the characteristic that two consecutive primary synchronization symbols are identical, as well as two consecutive secondary synchronization symbols, to obtain a frequency offset estimate using the autocorrelation method of the received signal, without needing to regenerate the reference sequence. The frequency offset is estimated by extracting the signal segment, calculating the conjugate cross-correlation phase difference, and then estimating the frequency offset.

[0089] ,

[0090] in That is, the time interval (in samples) between two repeated synchronization symbols. , and , These are the relevant segments of PSSS and secondary synchronization symbols (SSSS), respectively. The phase angle is then calculated. Next, based on fine compensation of the cyclic prefix, the 13 symbols are traversed to extract the CP segment and the data tail segment. The phase difference is calculated and averaged to obtain the high-precision frequency offset. The signal after phase rotation compensation :

[0091] ,

[0092] This represents the time interval between the CP and its sign tail; the starting position of the FFT is... After time-domain synchronization, the signal is The total frequency offset estimate This is the sum of the coarse and fine estimates, along with the sampling rate. .

[0093] Finally, for each symbol, the demodulation process first determines the FFT start position. , adopted as The calculation uses a factor of 0.55 to ensure the FFT starting point is located in the middle of the cyclic prefix, minimizing inter-symbol interference. During demodulation, a length of [length missing] is first extracted. The time-domain sample segment, and the factor used for half-subcarrier frequency offset compensation. Multiply the results, then perform an FFT transform on the compensated time-domain signal to convert it to the frequency domain. This is to correct for the factor... The phase distortion caused by the offset will be further affected by the phase compensation factor in the transformed frequency domain signal. Multiply them, and finally shift the zero-frequency component to the center. Among them... and Defined as:

[0094] ,

[0095] The index vector ranges from 0 to 2047. The method utilizes the structural characteristics of the cyclic prefix and eliminates the effects of time-domain offset and subcarrier misalignment through phase and half-subcarrier compensation to obtain frequency-domain resources.

[0096] Channel estimation of the preprocessed PSBCH signal is performed based on optimized LMMSE technology, specifically including:

[0097] First, estimate the channel power. The formula for characterizing signal energy is:

[0098] ,

[0099] Let be the number of subcarriers, where for The channel estimation flattening vector, when C is 2 or 3, corresponds to the PSSS symbol at the corresponding position, containing two columns of 62 subcarriers, for a total of 124 elements. To estimate the input noise power, the noise power is estimated by calculating the average energy of the noise segment. Subtract... Take the non-negative value last to ensure The result is not negative, providing channel energy parameters for subsequent correlation matrix construction. The least squares estimation result... Averaging across the symbol dimension and using the least-squares estimate of the averaged multi-symbols smooths noise and enhances channel response reliability. A zero-padding inverse fast Fourier transform converts the frequency-domain averaged channel response into the time-domain channel impulse response (CIR), representing the amplitude and delay information of each path in the multipath channel. The zero-frequency position of the CIR is adjusted to ensure time delay alignment. The power delay spectrum is calculated to provide the energy distribution of the multipath components. ,but:

[0100] ,

[0101] This reflects the channel delay characteristics and maintains consistency between the delay resolution and the frequency domain by eliminating redundant data introduced by zero-padding. For weak path interference, the signal-to-noise ratio is estimated. Set a dynamic threshold To focus on the main path:

[0102] ,

[0103] Convert the dB value to a linear scale and multiply by The maximum value is used to obtain the threshold, where, To ensure a minimum attenuation of 5dB, the threshold is prevented from being too high at low signal-to-noise ratios, thus preserving a large amount of noise components. The power delay distribution is then normalized to a probability distribution, facilitating the statistical analysis of delay characteristics. Assume the delay vector is... ,but:

[0104] ,

[0105] in It is the time-domain sampling period. For subcarrier spacing, RMS delay spread Defined as the root mean square deviation of the time delay, its calculation formula is:

[0106] ,

[0107] This is the average time delay, which is the expected time delay value of the power delay distribution. If the calculated result is invalid, the default value is used. Suitable for low signal-to-noise ratio scenarios;

[0108] Subsequently, the channel autocorrelation matrix The construction is based on root mean square delay spread, where The effective channel power is obtained by subtracting the noise power from the expected power value of the channel response using least squares estimation, and is finally defined as:

[0109] ,

[0110] The frequency difference between subcarriers, multiplied by The phase offset related to the time delay is obtained. It simulates the coherent bandwidth effect caused by the time delay and is applicable to multipath channels. To eliminate errors introduced by numerical calculations, a matrix... After adjusting for Hermitian symmetry, the original matrix is ​​added to its conjugate transpose and then divided by 2 to ensure that the matrix estimation maintains positive definiteness. Considering that channel estimation in vehicular networks is significantly affected by noise and multipath effects, especially under low signal-to-noise ratio or complex channel conditions, where estimation may degrade significantly, a correlation matrix-based approach is introduced. The obtained prior signal-to-noise ratio To optimize LMMSE estimation, The trace of the matrix, It integrates channel correlation and delay information to reflect more accurate channel statistical characteristics:

[0111] ,

[0112] And generate dynamically adjusted regularization terms. ,when When the frequency is high, reduce the impact of regularization and preserve channel details. When the regularization strength is low, the matrix stability is enhanced by increasing the regularization strength. Then, the frequency domain channel estimate is calculated using the following formula. :

[0113] ,

[0114] It is the identity matrix. The LMMSE weight matrix is ​​used to correct noise and system errors in the LS estimation using channel correlation, improving estimation accuracy. In multipath channels, CIR is usually concentrated on only a few samples in the time domain, which is related to RMS delay spread. In C-V2X, the typical value ranges from 0.1 to 5 microseconds, reflecting the short-lived nature of multipath paths in urban or highway environments. In contrast, since noise and radio frequency interference are widely distributed in the time domain, they do not have the concentration of CIR. The obtained frequency domain channel estimate can be transformed to the time domain using IDFT and then windowed to remove a wide range of noise and RFF, making the channel estimate more accurate. This is because noise components are suppressed while the main multipath paths of the channel are preserved. Finally, the time domain estimate after the operation is transformed to the frequency domain using DFT to obtain the frequency domain estimate. In low-speed or medium-speed vehicle scenarios, the channel transformation to adjacent symbols is small, so the frequency domain estimates of adjacent symbols can be approximated as the same. The frequency domain estimates of adjacent symbols are averaged, and the redundancy of the reference signal is used to smooth noise and interference to obtain the final frequency domain channel estimate.

[0115] Finally, to remove the influence of the wireless channel, channel information is removed through channel equalization technology, the initial RFF is extracted, and the received signal of the Cth frequency domain SC-FDMA symbol after preprocessing, CP removal, and DFT is represented as follows: The channel equalization operation is as follows:

[0116] ,

[0117] The initial RFF corresponding to the same data sequence can be further averaged and then used for identification in the proposed network;

[0118] Extracted fingerprints, such as Figure 4a and Figure 4b As shown;

[0119] The obtained radio frequency fingerprint data was trained using an improved ShuffleNet V2 network for cross-classification in different scenarios;

[0120] The network underwent targeted improvements, including expanding the channel and layer structure, introducing an attention mechanism, and implementing frequency domain signal optimization strategies. The overall architecture is as follows: Figure 5 As shown;

[0121] The Inverted Residual module is a core component of ShuffleNet V2, used to construct the various stages of the network. For example... Figure 6 As shown, when stride=1, channel splitting reduces computation while preserving identity mappings to support residual linking, such as... Figure 7As shown, when stride=2, the feature size is processed through two branches, and the number of channels is increased. Furthermore, considering that the extracted RFF signal is one-dimensional frequency domain data, to adapt to this characteristic, the network computing units are adjusted to a single-dimensional configuration to achieve feature extraction along the frequency domain axis, thereby enabling direct learning and modeling of the correlation between frequency points. Adjustments to the channel and layer structure, such as... Figure 5 As shown, Stage 2 has 4 blocks and 116 output channels, while Stage 3 has 8 blocks and 232 output channels. The deeper network structure enhances the ability to extract deep patterns from frequency domain signals, making it particularly suitable for capturing subtle spectral differences between devices. Channel expansion increases feature diversity, enabling the network to represent more potential frequency domain features. Stage 4 maintains 1024 channels to strike a balance between expressive power and computational cost.

[0122] A convolutional block attention module is integrated after the final convolutional layer. Through joint optimization of channel attention and spatial attention, key spectral features are dynamically focused. The formulas for channel and spatial attention are:

[0123] ,

[0124] The channel attention module extracts channel features through global average pooling and global max pooling, and generates channel weights by combining them with a fully connected layer. The spatial attention module generates spatial features on the feature map after channel attention weighting by average pooling and max pooling along the channel dimension, and then generates spatial weights by using one-dimensional convolution.

[0125] The channel shuffling operation of ShuffleNet V2 is retained, but mitigated through efficient feature fusion. Batch size, Number of groups For the number of channels, For signal length, the dimension implementation process during channel dimension shuffling is as follows:

[0126] ,

[0127] Finally, cosine annealing learning rate scheduling is used to optimize the training process, as shown in the formula:

[0128] ,

[0129] Initial learning rate, Minimum learning rate The total number of training rounds. For the current training round, For the first The specific learning rate value used in each round of model training;

[0130] During training, the model's generalization ability is enhanced by combining training datasets from different scenarios and using diverse samples.

[0131] The improved ShuffleNet V2 network employs channel shuffling operations and combines a convolutional neural network framework with an attention mechanism to construct a fingerprint classification model; further including:

[0132] Channel shuffling operations optimize computational efficiency through feature fusion. and Features after channel segmentation The channel splicing operation is defined as follows:

[0133] ,

[0134] The convolutional block attention module enhances the ability to extract key frequency offset features from RFF signals through joint optimization of channel and spatial dimensions.

[0135] The method in this embodiment achieves classification accuracy of 96.76% and 91.05% in static and dynamic scenes with low signal-to-noise ratios, respectively.

[0136] This embodiment is applicable to C-V2X communication scenarios, using 12 identical modules for transmission, with a total bandwidth of 20MHz, and equipped with a general-purpose software wireless point peripheral B205 with a carrier frequency of 5.9GHz and a sampling rate of 30.72Msps for receiving. Physical sidelink broadcast data with different signal-to-noise ratios in different scenarios are collected to form a dataset, which is then divided into training and test sets proportionally.

[0137] This embodiment also discloses a radio frequency fingerprint recognition system, including:

[0138] The signal receiving module is used to receive the PSBCH signal of the C-V2X system;

[0139] The signal preprocessing module is used to perform signal detection based on block energy gradient, coarse synchronization based on PSSS and fine synchronization based on CP, and CFO compensation based on synchronization signal and CP.

[0140] The channel estimation module uses a sample-optimized LMMSE algorithm to generate the frequency domain channel response and extract RFF features through multi-symbol LS estimation, RMS delay spread calculation, channel autocorrelation matrix construction, and time-domain lengthening operation.

[0141] The classification module uses an improved ShuffleNet V2 network, which performs device identification on the processed RFF features through expanded channel number, convolutional fast attention mechanism and learning rate scheduling.

[0142] The dataset integration module is used to combine training data from both visual and non-visual, static and dynamic scenes to optimize the model's generalization ability.

[0143] The method in this embodiment obtains the statistical characteristics of the channel response through the channel covariance matrix, considers the correlation between time and frequency, maintains the robustness of LMMSE, reduces excessive reliance on pilot data in high-noise environments, and suppresses noise amplification effects. Addressing the resource constraints of C-V2X, and based on an analysis of the model complexity in this field, a lightweight recognition network is constructed, exhibiting robustness to different scenarios and noise levels, and possessing the potential for on-site training and adjustment. It effectively considers the impact of noise and channel on fingerprints, enabling more accurate extraction of fingerprints from different devices in complex environments. It overcomes the limitations of traditional methods, effectively solving problems such as incomplete fingerprint extraction and classification accuracy in C-V2X systems.

[0144] This embodiment provides a method for radio frequency fingerprinting of vehicle-to-everything (V2X) broadcast frames based on LMMSE channel estimation, including the following steps:

[0145] Collect physical sidelink broadcast data with different signal-to-noise ratios in different scenarios to form a dataset, and divide the dataset into training set and test set according to the proportion;

[0146] The trained fingerprint model can be used to identify devices in other scenarios.

[0147] To verify the construction method of radio frequency fingerprinting and model, the identification method of this embodiment is compared with various traditional identification methods in the following experiments:

[0148] The data sources used are disclosed as follows:

[0149] For the experiment, 12 identical C-V2X modules were used to transmit PSBCH subframes, with a total bandwidth of 20MHz. A general-purpose software wireless point peripheral B205 with a carrier frequency of 5.9GHz and a sampling rate of 30.72Msps was used to receive the signals. Data acquisition environments included direct connection scenarios, stationary scenarios (LOS and NLOS), and moving scenarios (MOV1 (LOS), MOV2 (NLOS), and MOV3 (LOS+NLOS). In all these acquisitions, the transmitted PSBCHs were randomized, and each device in each scenario received approximately 1000 subframes. After preprocessing and fingerprint extraction, training and test datasets were constructed. All models were implemented using Python 3.12, and each model was trained for approximately 100 epochs. The model with the best results was selected for testing.

[0150] Furthermore, given the diversity of data acquisition scenarios and the balanced number of data packets per device, this embodiment uses overall accuracy as the primary evaluation metric to measure the proportion of correctly classified samples of device fingerprints in each scenario. When the class distribution is uniform, it is equivalent to macro-recall. In addition, the confusion matrix provides detailed prediction performance for each class, and the macro-F1 score, as the harmonic mean of precision and recall, evaluates the model's balanced performance across classes, as shown in the following formula:

[0151] ,

[0152] In the above formula, This indicates that for a specific category i, i.e., actually belonging to category i And it was correctly classified into categories by the model. The sample, It does not actually belong to any category. However, it was incorrectly classified into a category by the model. The sample, That is, it actually belongs to the category However, samples that were misclassified by the model as other categories, This indicates that it does not actually belong to the category. The total number of samples that are correctly predicted by the model as other categories.

[0153] Various recognition methods compared to the fingerprint recognition method of the present invention are divided into three categories: comparison of channel processing under different signal-to-noise ratios; classification under different scenarios; and comparison of different deep learning models.

[0154] (1) Classification experiments under different signal-to-noise ratios:

[0155] Figure 8 This demonstrates the trend of recognition accuracy variation within the SNR range of 0 dB to 30 dB. Figure 8 As can be seen, the accuracy of both methods increases with increasing SNR. In the low SNR region (0-10 dB), the LS method has lower accuracy, a lower initial value, and slower growth, indicating limited noise suppression capability. In contrast, the LMMSE method effectively suppresses noise interference under low SNR by utilizing channel autocorrelation and noise variance information. Even with SNR below 5 dB, its accuracy remains above 90%, and its growth is more stable, demonstrating its robustness advantage under low SNR conditions. When SNR reaches above 20 dB, the accuracy of both methods approaches 100%, but LMMSE maintains a slight advantage, highlighting its robustness under complex channel conditions.

[0156] Conduct cross-experiments in different scenarios:

[0157] The classification performance of the training and test sets under different scenarios is shown in Table 1, which presents the cross-authentication results of the fingerprints extracted after the proposed LMMSE channel estimation on test sets in multiple scenarios. High classification accuracy is achieved, with the highest accuracy being 100% and the lowest being 92.64%. Experimental results show that different training set combinations significantly affect the classification performance of the test set. Taking the direct connection scenario as an example, when the training set only contains direct connection data, the cross-authentication accuracy is 99.75%, 98.82%, and 99.01%, 98.09%, 99.59% in LOS, NLOS, and mobile scenarios 1-3, respectively. However, when the training set is expanded to a combination of direct connection and NLOS, the accuracy in the NLOS scenario reaches 100%, while the accuracy in the direct connection scenario drops to 95.64%. This may be attributed to the multipath effect in the NLOS scenario enhancing the fingerprint's discriminative power, while the channel balance in the direct connection scenario weakens feature diversity. Furthermore, cross-testing in some mobile scenarios performs better than in static scenarios. This is because in static scenarios, the fingerprints extracted after channel equalization lack some discriminative power due to the identical sampling locations and channels. However, overall, the combination of different training sets has a significant impact on classification performance.

[0158] Table 1 Performance comparison of fingerprint classification in different scenarios

[0159]

[0160] Model comparison experiment:

[0161] To evaluate the generalizability of the improved method to other deep learning models, a comparative analysis was conducted with other deep learning models, including DenseNet, MobileNet, EfficientNet, MobileBit, and ConvNeXt. Table 2 summarizes the performance of each model in mobile scenarios. The proposed model achieves an overall accuracy of 99.01%, and also performs well in macro-accuracy and macro-F1 score, outperforming other models. The improved channel estimation provides more accurate device fingerprint features to the model's input, and the introduction of a computationally inefficient convolutional block attention mechanism enhances the model's ability to focus on discriminative features, especially in the complex environment of mobile scenarios. In contrast, while DenseNet and MobileNet have advantages in computational efficiency, their lower accuracy and F1 score indicate limitations in handling multi-scenario fingerprint datasets. It is worth noting that most existing models rely on offline training, and the differences between offline and field data, such as device aging, can lead to poor performance in real-world applications. The lightweight nature of the proposed model makes on-site training and tuning possible.

[0162] Table 2 Comparison of fingerprint classification performance of different models

[0163] Deep learning models Overall accuracy (%) Macro Precision (%) Macro F1(%) DenseNet 95.19 93.15 95.58 MobileNet 92.7 88.93 88.12 EfficientNet 96.63 97.02 96.24 MobileVit 86.47 85.08 84.22 ConvNeXt 96.72 94 95.42 ShuffleNet V2 96.74 95.02 95.09 Model in this embodiment 99.01 97.44 98.05

[0164] After a series of verification analyses, the results show that the fingerprint recognition method in this embodiment has significant advantages in improving classification accuracy and coping with complex environments. Even when facing low signal-to-noise ratio conditions, fingerprints can still maintain high recognition accuracy.

[0165] An electronic device according to embodiments of this disclosure includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), a hard disk, flash memory, etc.

[0166] The processor may be a central processing unit (CPU) or other processing unit with data processing and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions. In one embodiment of this disclosure, the processor is used to run computer-readable instructions stored in the memory, causing the electronic device to perform all or part of the steps of the LMMSE channel estimation-based vehicular network broadcast frame radio frequency fingerprinting method of the foregoing embodiments of this disclosure.

[0167] Those skilled in the art will understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.

[0168] The electronic device provided in this disclosure may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) or a program loaded from a storage device into a random access memory (RAM). The RAM also stores various programs and data required for the operation of the electronic device. The processing device, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0169] Typically, the following devices can be connected to the I / O interface: input devices, such as sensors or visual information acquisition devices; output devices, such as displays; storage devices, such as magnetic tapes or hard drives; and communication devices. Communication devices allow electronic devices to communicate wirelessly or wiredly with other devices (such as edge computing devices) to exchange data.

[0170] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processing device, all or part of the steps of the vehicular network broadcast frame radio frequency fingerprinting method based on LMMSE channel estimation according to embodiments of this disclosure are performed.

[0171] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0172] A computer-readable storage medium according to embodiments of the present disclosure stores non-transitory computer-readable instructions. When these non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the vehicular network broadcast frame radio frequency fingerprinting method based on LMMSE channel estimation according to the foregoing embodiments of the present disclosure are performed.

[0173] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).

[0174] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0175] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0176] In this disclosure, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The block diagrams of devices, apparatuses, devices, and systems involved in this disclosure are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as "comprising," "including," "having," etc., are open-ended terms meaning "including but not limited to," and are used interchangeably with them. The terms "or" and "and" as used herein refer to the terms "and / or," and are used interchangeably with them unless the context clearly indicates otherwise. The term "such as" as used herein refers to the phrase "such as but not limited to," and is used interchangeably with it.

[0177] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.

[0178] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.

[0179] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.

[0180] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0181] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A method for radio frequency fingerprinting of vehicle-to-everything (V2X) broadcast frames based on LMMSE channel estimation, characterized in that, include: Subcarrier data in the resource grid is obtained based on the signal from the acquired physical sidelink broadcast channel; Based on subcarrier data, root mean square delay spread and channel autocorrelation matrix construction are performed. Combined with prior signal-to-noise ratio regularization and time-domain windowing operation, the channel estimate is obtained. Initial radio frequency fingerprint features are obtained through channel equalization and channel estimation, and vehicle network devices are classified and identified using an improved neural network and initial radio frequency fingerprint features. The process of performing root mean square delay spread and constructing the channel autocorrelation matrix based on subcarrier data, combined with prior signal-to-noise ratio regularization and time-domain windowing, yields the channel estimate, including: Estimate channel power to characterize signal energy; The channel autocorrelation matrix is ​​constructed based on the estimated channel power and root mean square delay spread. The channel autocorrelation matrix is ​​constructed based on the estimated channel power and the delay-related phase offset, which is obtained based on the frequency difference between subcarriers. By combining prior signal-to-noise ratio regularization and time-domain windowing, channel estimates are obtained, including: Generate dynamically adjusted regularization terms Regularization term It is used to reduce the impact of regularization when the prior signal-to-noise ratio is high, and to increase the regularization strength when the prior signal-to-noise ratio is low. The process of performing root mean square delay spread and constructing the channel autocorrelation matrix based on subcarrier data, combined with prior signal-to-noise ratio regularization and time-domain windowing, yields the channel estimate, including: In low signal-to-noise ratio scenarios, the default setting is the root mean square delay spread value; The initial frequency domain channel response is averaged using multiple symbols to smooth out the noise effects. The data differences in the channel autocorrelation matrix are eliminated by symmetry processing, and the main path is preserved by windowing operation to suppress noise and radio frequency fingerprint interference. Initial RF fingerprint features are obtained through channel equalization and channel estimation. The improved neural network used in classifying and identifying vehicular network devices based on these initial RF fingerprint features is an improved ShuffleNet V2 network. The improved ShuffleNet V2 network includes: A convolutional block attention module is integrated after the final convolutional layer, and key spectral features are dynamically focused through joint optimization of channel attention and spatial attention. The channel attention module extracts channel features through global average pooling and global max pooling, and generates channel weights by combining them with a fully connected layer. The spatial attention module generates spatial features on the feature map after channel attention weighting by average pooling and max pooling along the channel dimension, and then generates spatial weights by using one-dimensional convolution.

2. The vehicular network broadcast frame radio frequency fingerprinting method based on LMMSE channel estimation according to claim 1, characterized in that, The subcarrier data in the resource grid is obtained based on the signal of the acquired physical side link broadcast channel, including: signal detection, frame synchronization, carrier frequency offset compensation and resource grid demodulation; The signal detection method uses block energy gradient and threshold determination to locate the signal starting point by calculating the energy ratio of adjacent data blocks. The frame synchronization utilizes the periodic characteristics of the master synchronization signal for coarse synchronization and searches for signal peak points through amplitude normalization cross-correlation calculation. The carrier frequency offset compensation utilizes the characteristic that two consecutive primary synchronization symbols are the same and two consecutive secondary synchronization symbols are the same, and obtains the frequency offset estimate by using the signal autocorrelation of the physical side link broadcast channel. The resource grid demodulation utilizes the structural characteristics of the cyclic prefix and eliminates the effects of time domain offset and subcarrier misalignment through phase and half-subcarrier compensation to obtain frequency domain resources.

3. The vehicular network broadcast frame radio frequency fingerprinting method based on LMMSE channel estimation according to claim 2, characterized in that, The signal detection includes: Define the signal block length as Signals that traverse the physical side link broadcast channel According to block length Divide the data into non-overlapping segments and calculate the energy block by block. : ， in For the first Grid data blocks, For sampling point index, for transpose, It is a natural number.

4. The vehicular network broadcast frame radio frequency fingerprinting method based on LMMSE channel estimation according to claim 3, characterized in that, The carrier frequency offset compensation includes: Extract the signal segment, calculate the conjugate cross-correlation phase difference, and estimate the frequency offset: ； in, For frequency offset, , The number of points in the Fourier transform. and The relevant segment of the main synchronization signal and The relevant segment for the auxiliary synchronization symbol, where n is a natural number.

5. The vehicular network broadcast frame radio frequency fingerprinting method based on LMMSE channel estimation according to claim 4, characterized in that, The resource grid demodulation includes: Determine the starting position of the Fourier transform, and then extract the length as... The time-domain sample segment, with a length of Time-domain sample segments and factors Multiply, then perform a Fourier transform on the multiplied time-domain signal to convert the time-domain sample segment to the frequency domain, factoring... Used for half-subcarrier frequency offset compensation.

Images