Cross-channel radio frequency fingerprint identification method based on confidence perception and dynamic fusion

Abstract

The invention discloses a cross-channel radio frequency fingerprint identification method based on confidence perception and dynamic fusion, and belongs to the technical field of wireless communication and signal processing. The method comprises the following steps: acquiring an original IQ signal and generating a corresponding equalization signal, preprocessing the original signal and the equalization signal, constructing a data set with a transmitter label, and dividing the data set into a training set, a verification set and a test set; constructing a double-branch feature extraction network comprising an original signal branch and an equalization signal branch, and introducing a cross-branch interaction module; and training the network by using training data, and dynamically fusing the two branch features based on a confidence perception mechanism in a test stage to realize radio frequency fingerprint identification. By adaptively fusing the characteristics of the original signal and the balanced signal, the accuracy and robustness of radio frequency fingerprint identification in a cross-channel scene are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of wireless communication and signal processing technology, specifically to a cross-channel radio frequency fingerprinting method based on confidence perception and dynamic fusion. Background Technology

[0002] Radio frequency fingerprinting (RFF) technology leverages minute manufacturing differences in transmitter hardware to identify and authenticate wireless devices. Compared to traditional cryptographic methods, RFF offers significant advantages such as lightweight nature, difficulty in cloning, and the elimination of the need for additional protocol interactions, making it an important research area in physical layer security.

[0003] In cross-channel applications, deep learning-based RF fingerprinting methods generally suffer from performance degradation. This is primarily because different channel conditions cause multi-dimensional distortions in signal amplitude, phase, and frequency, thus masking the fine-grained features resulting from differences in RF front-end hardware.

[0004] Existing technologies typically employ channel equalization to reduce the aforementioned distortions and obtain a more stable signal representation. However, the equalization process can cause a smoothing effect on the subtle structure of the signal, which is detrimental to preserving the inherent RF fingerprint characteristics of the device. This makes it difficult for identification methods relying solely on the equalized signal to achieve the desired accuracy. Therefore, using only the original signal or only the equalized signal is insufficient to simultaneously ensure both channel interference resistance and fingerprint feature fidelity. How to suppress channel influences while avoiding fingerprint feature loss, and how to achieve effective complementary utilization between the original and equalized signals, has become a pressing technical problem that needs to be solved.

[0005] To address this, the present invention proposes a cross-channel radio frequency fingerprinting method based on confidence perception and dynamic fusion. Summary of the Invention

[0006] The purpose of this invention is to provide a cross-channel radio frequency fingerprinting method based on confidence perception and dynamic fusion. By designing a dual-branch feature extraction network and a cross-branch interaction module, combined with a confidence-based dynamic fusion mechanism, adaptive fusion of the original signal and the equalization signal is achieved, which effectively improves the recognition performance and robustness in cross-channel scenarios.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a cross-channel radio frequency fingerprinting method based on confidence perception and dynamic fusion, comprising the following steps: The raw IQ signal is acquired and the corresponding equalization signal is generated. The raw IQ signal and the equalization signal are preprocessed to construct a labeled dataset containing transmitter labels. The dataset is divided into training set, validation set and test set. A dual-branch feature extraction network is constructed, consisting of an original signal branch and an equalized signal branch. Each branch contains a patch embedding layer and a multilayer perceptron layer for extracting local and global feature representations. The dual-branch feature extraction network was trained and tested using training and testing sets, and the trained dual-branch feature extraction network was used to realize radio frequency fingerprint recognition in cross-channel scenarios. In particular, a cross-branch interaction module is introduced during the training phase of the dual-branch feature extraction network. This module is used to randomly exchange feature representations between the original signal branch and the equalized signal branch, and to construct a composite loss function to drive the dual-branch feature extraction network to achieve end-to-end collaborative optimization of feature extraction, confidence estimation and adaptive fusion in cross-channel scenarios. During the testing phase of the dual-branch feature extraction network, a dynamic fusion mechanism is constructed. The confidence levels of the original signal branch and the equalization signal branch are evaluated in real time through the energy function, and adaptive weighted fusion is performed to achieve adaptive fusion of the original signal and the equalization signal to enhance the recognition robustness.

[0008] Furthermore, the original IQ signals are acquired and corresponding equalization signals are generated. The original IQ signals and equalization signals are preprocessed to construct a labeled dataset containing transmitter labels, as follows: (21) In multi-receiver and cross-channel scenarios, the original IQ sampling sequence of the wireless transmitting device is collected. Then, channel state information is obtained using pilot signals, training sequences, or channel estimation algorithms. Channel equalization processing is performed on the original signal to generate an equalized signal that is time-domain aligned and corresponds one-to-one with the original signal. Thus, a dual-mode dataset containing the original signal and the equalized signal is established. The original signal and the equalized signal are represented as follows: in The number of sampling points. This indicates the real and imaginary parts of the IQ signal; the device label is... , Number of equipment categories; (22) Perform uniform standardization operations on the original signal and the equalized signal, including amplitude and phase normalization, fixed-length truncation or zero padding, and outlier sample removal; on this basis, associate device identity tags and environment tags with each signal pair, and finally construct a labeled cross-channel dataset with consistent format and containing dual-branch input.

[0009] Furthermore, the patch embedding layer is implemented using a one-dimensional convolutional layer, with a kernel size of Spatch and a stride of [missing value]. ; The patch embedding layer will embed the original signal and equalization signal The process involves segmenting the data into overlapping local patches and projecting them onto a high-dimensional embedding space using a one-dimensional convolutional layer. This process can be represented as follows: in For the number of patches, For the embedded dimension; This process is used to preserve the local timing structure and amplitude-phase correlation of the IQ signal.

[0010] Furthermore, each multilayer perceptron layer comprises a layer normalized network (LN), a feedforward network (FFN), and residual connections, specifically represented as follows: Here, FFN is defined as a two-layer fully connected network containing the GELU activation function, and is defined as follows: in, For GELU activation function, and This is a learnable weight matrix used to expand the feature dimension from D to rD and then map it back to D, where r is a predefined expansion ratio. and This is a learnable bias term.

[0011] Furthermore, the patch is embedded into the sequence input multilayer perceptron layer for feature refinement. Global average pooling (GAP) is performed on the refined features of each branch, and then a fully connected classification head f is used. cls Generate predictions Normalization Generate prediction probability , is represented as: in These are the features after the exchange.

[0012] Final fusion prediction probability Represented as: in and The fusion weights are adaptively optimized for a dynamic fusion mechanism.

[0013] Furthermore, a cross-branch interaction module is introduced during the training phase of the dual-branch feature extraction network to randomly exchange feature representations between the original signal branch and the equalized signal branch, as detailed below: The cross-branch interaction module generates random variables that follow a uniform distribution. To control the random exchange of features, when At this time, a feature exchange operation is performed between the two branches, that is, the original signal branch receives the features of the equalized signal branch, and the equalized signal branch receives the features of the original signal branch; when At this time, the original feature allocation remains unchanged, and this process can be expressed as: In the formula, For equalizing signals, The original signal; Furthermore, the dynamic fusion mechanism is constructed by evaluating the confidence levels of the original signal branch and the equalized signal branch in real time through an energy function, and then performing adaptive weighted fusion, as follows: (71) Based on the outputs of the original signal branch and the equalization signal branch Calculate the energy fraction, which serves as a metric for measuring the uncertainty of branch predictions. Energy of each branch Defined as: in For the first Branch in the On class Output; (72) Calculate fusion weights based on energy fractions using linear mapping , is represented as: In the formula, the slope is set. And bias ,because This formula ensures that high-energy, low-confidence branches are assigned smaller weights, thereby achieving confidence-oriented adaptive fusion. Then it was substituted into the formula The final prediction was completed in the middle; (73) Definition For the first The sample at the th Historical average training loss on each branch: in, The starting statistical round, Epoch. The total number of sampling rounds, For real labels, The parameter is Branching model at time, Cross-entropy loss; The larger the value, the more difficult it is for the model to classify the sample correctly in the historical training, i.e., it belongs to the "difficult sample"; otherwise, it is the "easy sample".

[0014] Furthermore, a composite loss function is constructed to drive the dual-branch feature extraction network to achieve end-to-end collaborative optimization of feature extraction, confidence estimation, and adaptive fusion in cross-channel scenarios, as detailed below: (81) To show the inverse correlation between enhancement energy and fusion weight, a ranking regularization term is introduced, expressed as: in and This represents two different samples from the same branch. It is a symbolic function; The core logic of this formula lies in: if the sample Historical average loss Larger than the sample of Smaller fusion weights Otherwise, a penalty loss will occur, thereby strengthening the inverse correlation between energy and weight, and thus achieving stable fusion; (82) Construct a composite total loss function that includes the main classification loss, the auxiliary branch loss, and the ranking regularization term. : in, The first term represents the cross-entropy loss function. The second term is used to supervise the final prediction results after fusion. The third term is used to separately supervise the independent predictive capabilities of the original and equilibrium branches. The logic used to generate constraint weights.

[0015] Furthermore, during the training phase of the dual-branch feature extraction network, a composite total loss including cross-entropy classification loss and ranking regularization term is calculated. The Adam optimizer is used to update the network parameters, forcing the model to perform cross-branch collaborative learning in the feature space, while establishing a high-precision confidence estimation mapping until the model converges.

[0016] Furthermore, during the testing phase of the dual-branch feature extraction network, the cross-branch interaction module was removed, allowing the two branches to process their respective input signals independently. For each test sample to be identified, the dynamic inference process based on energy confidence is executed as follows: Acquisition: Independent output of the original branch in a two-branch network With equilibrium branch ; Confidence assessment: Real-time calculation of energy fraction for each branch and Quantify the prediction uncertainty of the current sample under each branch; Dynamic weighting: utilizing pre-trained linear mapping relationships Generate instance-level fusion weights and ; Decision output: Calculate the final probability distribution after weighted fusion. The final device prediction label is determined using the maximum a posteriori probability criterion. : In the formula, This indicates the device's predicted label.

[0017] This invention has at least the following beneficial effects: 1. This invention achieves adaptive weighting of the original signal branch and the equalized signal branch through a dynamic fusion mechanism. It can evaluate the prediction confidence of each branch in real time and automatically adjust the fusion weights based on current channel conditions and sample characteristics. Compared to single-source methods, this mechanism effectively overcomes the polarization defects of the original signal being susceptible to channel distortion interference and the equalized signal easily losing fingerprint details. While suppressing channel influences, it maximizes the preservation of the fine-grained structure of the device hardware features, thus maintaining stable and high-precision recognition capabilities under different channel environments and receiving device conditions, significantly enhancing cross-channel generalization performance.

[0018] 2. In this invention, the cross-branch interaction module forces the dual-branch network to learn more complementary feature representations through a random feature exchange mechanism during the training phase, breaking the semantic isolation between the feature spaces of the original signal and the equalization signal. This design enables the network to mine the intrinsic correlation between the two types of signals in the device identity representation, enhances the model's ability to jointly model channel distortion and hardware fingerprint differences, and makes the dual-branch feature extraction process no longer a simple parallel computation, but a collaborative learning paradigm that promotes mutual development and information sharing, thereby improving the overall feature discrimination power.

[0019] 3. This invention introduces a ranking regularization term based on historical training loss statistics. It utilizes the "easy to difficult" fitting characteristic of samples during the training process to construct a quantitative evaluation benchmark for difficult samples. It explicitly constrains the inverse correlation between energy scores and fusion weights, effectively avoiding the fluctuations and biases that may occur in traditional confidence estimation based on a single energy value. This makes the fusion weight allocation mechanism more closely match the true difficulty distribution of samples, significantly improving the stability and reliability of confidence assessment.

[0020] 4. In this invention, the design of the composite loss function organically integrates the main classification loss, auxiliary branch loss, and ranking regularization term, driving the three major modules of feature extraction, confidence evaluation, and dynamic fusion to achieve end-to-end joint optimization under a unified framework. This avoids the local optimum problem that may be caused by phased training. Furthermore, by removing the cross-branch interaction module and solidifying the network topology during the inference phase, the determinism and latency stability of the system output are ensured, meeting the strict requirements of consistency and real-time performance in actual authentication scenarios.

[0021] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the method described in this invention; Figure 2 This is a structural diagram of the dual-branch feature embedding and cross-branch interaction module of the present invention; Figure 3 This is a schematic diagram illustrating the principle of the dynamic fusion mechanism based on energy confidence in this invention. Detailed Implementation

[0023] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0024] Please see Figures 1-3 This invention provides a technical solution: a cross-channel radio frequency fingerprint recognition method based on confidence perception and dynamic fusion, comprising the following steps: Step 1: Acquire raw IQ signals and generate corresponding equalization signals to construct a labeled dataset, with each sample containing a transmitter label. Normalize the signals and partition the data to form training, validation, and test sets.

[0025] Step 2: Construct a network structure with two feature extraction branches: the first branch processes the original signal, and the second branch processes the equalized signal. Each branch contains a patch embedding layer and a multilayer perceptron (MLP) layer for extracting local and global feature representations. During the training phase, a cross-branch interaction (CBI) module is introduced to randomly exchange feature representations between the two branches to enhance feature complementarity and cross-channel generalization ability.

[0026] Step 3: Construct a dynamic fusion (DF) mechanism, which evaluates the confidence level of each branch in real time through an energy function and adaptively adjusts the fusion weights to achieve confidence-aware branch fusion.

[0027] Step 4: Implement the complete training and inference process based on CDFNet. During the training phase, a composite loss function including a ranking regularization term is constructed to drive the dual-branch network to perform end-to-end feature co-learning and confidence mapping calibration. During the testing phase, the network topology is reconstructed to ensure system determinism, and a dynamic fusion mechanism based on energy functions is used to achieve instance-level adaptive weighting and highly robust identification of cross-channel samples.

[0028] In a further embodiment, step 1 involves acquiring and constructing a labeled cross-channel wireless signal dataset, specifically including the following steps: Step 1-1: In multi-receiver and cross-channel scenarios, acquire the original IQ sampling sequence of the wireless transmitting device; subsequently, obtain channel state information using pilot signals, training sequences, or channel estimation algorithms, perform channel equalization processing on the original signal, and generate an equalized signal that is time-domain aligned and corresponds one-to-one with the original signal, thereby establishing a signal set containing both "original" and "equalized" perspectives. The original signal and the equalized signal are represented as follows: in The number of sampling points. This represents the I and Q signals (real and imaginary parts). The device label is... , Number of equipment categories; Steps 1-2: Perform uniform standardization operations on the original signal and the equalized signal, including amplitude and phase normalization, fixed-length truncation or zero padding, and outlier sample removal; on this basis, synchronously associate each signal pair with device identity tags and environment tags, and finally construct a labeled cross-channel dataset with consistent format and containing dual-branch input, providing a standardized data foundation for subsequent feature extraction and dynamic fusion. The core of this step is to construct a dual-modal dataset containing both raw and equalized signals. By performing unified preprocessing, we ensure that the two signals establish a strict one-to-one correspondence and alignment relationship in terms of device and domain dimensions, providing an accurate data benchmark for subsequent feature fusion. In a further embodiment, step 2 constructs a two-branch network with a cross-branch interaction (CBI) mechanism, specifically including the following steps: Step 2-1: In order to preserve the local timing structure and amplitude-phase correlation of the IQ signal, firstly, the input signal... and The data is segmented into overlapping local patches and projected onto a high-dimensional embedding space through a one-dimensional convolutional layer. The kernel size is Spatch, and the stride is [missing value]. The process is represented as: in For the number of patches, As an embedding dimension, this process preserves the local temporal structure and amplitude-phase correlation of the IQ signal, providing a compact signal representation for subsequent modules; Step 2-2: Input the patch embedding sequence into the MLP module for feature refinement. Each MLP module contains a layer normalization (LN), a feedforward network (FFN), and residual connections, represented as follows: Here, FFN is defined as a two-layer fully connected network containing the GELU activation function, and is defined as follows: in, For GELU activation function, and This is a learnable weight matrix used to expand the feature dimension from D to rD and then map it back to D, where r is a predefined expansion ratio. and For learnable bias terms; Steps 2-3: To enhance the complementarity between the two branches, a cross-branch interaction (CBI) module is introduced during the training phase. This module generates random variables that follow a uniform distribution. To control the random exchange of features, when At that time, a feature exchange operation is performed between the two branches, that is, the original signal branch receives the features of the equalized signal branch. The characteristics of the equalized signal branch receiving the original signal branch ;when At this time, the original feature allocation remains unchanged, and this process can be expressed as: In the formula, For equalizing signals, The original signal; Steps 2-4: Perform global average pooling on the refined features of each branch. And through fully connected classification heads Generate predictions Normalization Generate prediction probability , is represented as: In the formula, Features after the exchange; Final fusion prediction probability Represented as: in and The fusion weights are then adaptively optimized using the DF mechanism. In a further embodiment, step 3 constructs a dynamic fusion (DF) mechanism based on the energy function to achieve branch fusion with dynamic confidence perception. The specific steps include: Step 3-1: Based on the output of each branch Calculate the energy score, which serves as a metric for measuring the uncertainty of branch predictions. Energy of each branch Defined as: in For the first Branch in the On class Output; Step 3-2: Calculate the fusion weights based on the energy scores using a linear mapping, expressed as: In this embodiment, the slope is set. And bias ; because This formula ensures that high-energy (low-confidence) branches are assigned smaller weights, thereby achieving confidence-oriented adaptive fusion. It is then substituted into the formula described in step 2. The final prediction was completed in the middle; Step 3-3: To address the issue of inaccurate confidence estimation, the model's historical learning trajectory is used to quantify the sample difficulty. Based on the neural network's "easy-to-difficult" fitting characteristic, a definition is defined. For the first The sample at the th Historical average training loss on each branch: in, This is the starting statistical round (Epoch). The total number of sampling rounds, For real labels, The parameter is Branching model at time, This represents the cross-entropy loss. The larger the value, the more difficult the sample has been to be correctly classified by the model during historical training, i.e., it belongs to the "hard sample" category; conversely, a smaller value indicates a "easy sample". This metric provides a stable "difficulty benchmark" for subsequent regularization constraints; Steps 3-4: To show the inverse correlation between enhancement energy and fusion weights, a sorting regularization term is introduced, expressed as: in and This represents two different samples from the same branch. For the sign function. The core logic of this formula is: if the sample... Historical average loss Larger than the sample of (Right now Compare (More difficult), then the model must be given Smaller fusion weights Otherwise, a penalty loss will occur. This strengthens the inverse correlation between energy and weights, thereby achieving a more stable and reliable fusion. Steps 3-5: To improve recognition accuracy while ensuring the interpretability and stability of the fusion weights, a composite total loss function is constructed, which includes the main classification loss, auxiliary branch loss, and ranking regularization term. : in, The first term represents the cross-entropy loss function. The second term is used to supervise the final prediction results after fusion. The third term is used to separately supervise the independent predictive capabilities of the original and equilibrium branches. The logic used to generate constraint weights; By minimizing this total loss, the model can achieve end-to-end collaborative optimization of feature extraction, confidence estimation, and adaptive fusion in cross-channel scenarios; In a further embodiment, step 4 executes an end-to-end model optimization training and cross-channel inference strategy based on CDFNet. This step clarifies the parameter update mechanism of the network during the training phase and the dynamic inference logic during the testing phase. Specific steps include: Step 4-1: During the training phase, construct a two-branch network embedding a cross-branch interaction (CBI) module, employ an end-to-end iterative optimization strategy, and calculate the composite total loss including cross-entropy classification loss and ranking regularization term. The network parameters are updated using the Adam optimizer, a process designed to force the model to perform cross-branch collaborative learning in the feature space while establishing a high-precision confidence estimation mapping until the model converges. Step 4-2: During the testing phase, the CBI module is removed to ensure the determinism and stability of the inference process. This operation aims to eliminate the uncertainty caused by random feature swapping, ensuring that the two branches independently process their respective input signals during inference, thus guaranteeing the stability and determinism of the model output and making it suitable for practical authentication scenarios with high requirements for latency and consistency. For each test sample to be identified, the dynamic inference process based on energy confidence is executed as follows: 1. Acquisition: Independent output of the original branch in a two-branch network With equilibrium branch ; 2. Confidence assessment: Real-time calculation of energy scores for each branch. and Quantify the prediction uncertainty of the current sample under each branch; 3. Dynamic weighting: Utilizing pre-trained linear mapping relationships Generate instance-level fusion weights and ; 4. Decision Output: Calculate the final probability distribution after weighted fusion. The final device prediction label is determined using the maximum a posteriori probability criterion. : In summary, this invention achieves adaptive weighting of the original signal branch and the equalized signal branch through a dynamic fusion mechanism. It can evaluate the prediction confidence of each branch in real time and automatically adjust the fusion weights based on current channel conditions and sample characteristics. Compared to single-source methods, this mechanism effectively overcomes the polarization defects of the original signal being susceptible to channel distortion interference and the equalized signal easily losing fingerprint details. While suppressing channel influences, it maximizes the preservation of the fine-grained structure of the device hardware features, thus maintaining stable and high-precision recognition capabilities under different channel environments and receiving device conditions, significantly enhancing cross-channel generalization performance.

[0029] It should be noted that, in this document, relational terms such as "first" and "second" are used only 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. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0030] For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances. When an element is referred to as being "assembled on," "mounted on," "fixed to," or "set on" another element, it may be directly on the other element or there may be an intermediate element present. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be an intermediate element present. The terms "vertical," "horizontal," "upper," "lower," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible embodiments.

[0031] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0032] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

Claims

1. A cross-channel radio frequency fingerprint recognition method based on confidence perception and dynamic fusion, characterized in that, Includes the following steps: The raw IQ signal is acquired and the corresponding equalization signal is generated. The raw IQ signal and the equalization signal are preprocessed to construct a labeled dataset containing transmitter labels. The dataset is divided into training set, validation set and test set. A dual-branch feature extraction network is constructed, consisting of an original signal branch and an equalized signal branch. Each branch contains a patch embedding layer and a multilayer perceptron layer for extracting local and global feature representations. The dual-branch feature extraction network was trained and tested using training and testing sets, and the trained dual-branch feature extraction network was used to realize radio frequency fingerprint recognition in cross-channel scenarios. In the training phase of the dual-branch feature extraction network, a cross-branch interaction module is introduced to randomly exchange feature representations between the original signal branch and the equalized signal branch, and to construct a composite loss function to drive the dual-branch feature extraction network to achieve end-to-end collaborative optimization of feature extraction, confidence estimation and adaptive fusion in cross-channel scenarios. During the testing phase of the dual-branch feature extraction network, a dynamic fusion mechanism is constructed. The confidence levels of the original signal branch and the equalization signal branch are evaluated in real time through the energy function, and adaptive weighted fusion is performed to achieve adaptive fusion of the original signal and the equalization signal to enhance the recognition robustness.

2. The cross-channel radio frequency fingerprint recognition method based on confidence perception and dynamic fusion according to claim 1, characterized in that: Raw IQ signals are acquired and corresponding equalization signals are generated. The raw IQ signals and equalization signals are preprocessed to construct a labeled dataset containing transmitter labels, as detailed below: (21) In multi-receiver and cross-channel scenarios, the original IQ sampling sequence of the wireless transmitting device is collected. Then, channel state information is obtained using pilot signals, training sequences, or channel estimation algorithms. Channel equalization processing is performed on the original signal to generate an equalized signal that is time-domain aligned and corresponds one-to-one with the original signal. Thus, a dual-mode dataset containing the original signal and the equalized signal is established. The original signal and the equalized signal are represented as follows: in The number of sampling points. This indicates the real and imaginary parts of the IQ signal; the device label is... , Number of equipment categories; (22) Perform uniform standardization operations on the original signal and the equalized signal, including amplitude and phase normalization, fixed-length truncation or zero padding, and outlier removal. Based on this, each signal pair is synchronously associated with a device identity label and an environment label, ultimately constructing a labeled cross-channel dataset with a consistent format and containing dual-branch inputs.

3. The cross-channel radio frequency fingerprint recognition method based on confidence perception and dynamic fusion according to claim 1, characterized in that: The patch embedding layer is implemented using a one-dimensional convolutional layer, with a kernel size of Spatch and a stride of [missing value]. ; The patch embedding layer will embed the original signal and equalization signal The process involves segmenting the data into overlapping local patches and projecting them onto a high-dimensional embedding space using a one-dimensional convolutional layer. This process can be represented as follows: in For the number of patches, For the embedded dimension; This process is used to preserve the local timing structure and amplitude-phase correlation of the IQ signal.

4. The cross-channel radio frequency fingerprint recognition method based on confidence perception and dynamic fusion according to claim 3, characterized in that: Each multilayer perceptron layer consists of a layer normalized network (LN), a feedforward network (FFN), and residual connections, specifically represented as follows: Here, FFN is defined as a two-layer fully connected network containing the GELU activation function, and is defined as follows: in, For GELU activation function, and This is a learnable weight matrix used to expand the feature dimension from D to rD and then map it back to D, where r is a predefined expansion ratio. and This is a learnable bias term.

5. The cross-channel radio frequency fingerprint recognition method based on confidence perception and dynamic fusion according to claim 4, characterized in that: The patch embedding sequence is input into a multilayer perceptron layer for feature refinement. Global average pooling (GAP) is performed on the refined features of each branch, and then the features are passed through a fully connected classification head f. cls Generate predictions Normalization Generate prediction probability , is represented as: in These are the features after the exchange. Final fusion prediction probability Represented as: in and The fusion weights are adaptively optimized for a dynamic fusion mechanism.

6. The cross-channel radio frequency fingerprint recognition method based on confidence perception and dynamic fusion according to claim 5, characterized in that: The introduction of a cross-branch interaction module during the training phase of the dual-branch feature extraction network is used to randomly exchange feature representations between the original signal branch and the equalized signal branch, as detailed below: The cross-branch interaction module generates random variables that follow a uniform distribution. To control the random exchange of features, when At this time, a feature exchange operation is performed between the two branches, that is, the original signal branch receives the features of the equalized signal branch, and the equalized signal branch receives the features of the original signal branch; when At this time, the original feature allocation remains unchanged, and this process can be expressed as: in For equalizing signals, This is the original signal.

7. The cross-channel radio frequency fingerprint recognition method based on confidence perception and dynamic fusion according to claim 6, characterized in that: The aforementioned dynamic fusion mechanism assesses the confidence levels of the original signal branch and the equalized signal branch in real time using an energy function, and performs adaptive weighted fusion, as detailed below: (71) Based on the outputs of the original signal branch and the equalization signal branch Calculate the energy fraction, which serves as a metric for measuring the uncertainty of branch predictions. Energy of each branch Defined as: in For the first Branch in the On class Output; (72) Calculate fusion weights based on energy fractions using linear mapping , is represented as: In the formula, the slope is set. And bias ,because This formula ensures that high-energy, low-confidence branches are assigned smaller weights, thereby achieving confidence-oriented adaptive fusion. Then it was substituted into the formula The final prediction was completed in the middle; (73) Definition For the first The sample at the th Historical average training loss on each branch: in, The starting statistical round, Epoch. The total number of sampling rounds, For real labels, The parameter is Branching model at time, Cross-entropy loss; The larger the value, the more difficult it is for the model to classify the sample correctly in the historical training, i.e., it belongs to the "difficult sample"; otherwise, it is the "easy sample".

8. The cross-channel radio frequency fingerprint recognition method based on confidence perception and dynamic fusion according to claim 6, characterized in that: A composite loss function is constructed to drive the dual-branch feature extraction network to achieve end-to-end collaborative optimization of feature extraction, confidence estimation, and adaptive fusion in cross-channel scenarios, as detailed below: (81) To show the inverse correlation between enhancement energy and fusion weight, a ranking regularization term is introduced, expressed as: in and This represents two different samples from the same branch. It is a symbolic function; The core logic of this formula lies in: if the sample Historical average loss Larger than the sample of Then the model must be given Smaller fusion weights Otherwise, a penalty loss will occur, thereby strengthening the inverse correlation between energy and weight, and thus achieving stable fusion; (82) Construct a composite total loss function that includes the main classification loss, the auxiliary branch loss, and the ranking regularization term. : in, The first term represents the cross-entropy loss function. The second term is used to supervise the final prediction results after fusion. The third term is used to separately supervise the independent predictive capabilities of the original and equilibrium branches. The logic used to generate constraint weights.

9. The cross-channel radio frequency fingerprint recognition method based on confidence perception and dynamic fusion according to claim 8, characterized in that: During the training phase of the two-branch feature extraction network, the composite total loss, including cross-entropy classification loss and ranking regularization term, is calculated. The Adam optimizer is used to update the network parameters, forcing the model to perform cross-branch collaborative learning in the feature space, while establishing a high-precision confidence estimation mapping until the model converges.

10. The cross-channel radio frequency fingerprint recognition method based on confidence perception and dynamic fusion according to claim 9, characterized in that: During the testing phase of the dual-branch feature extraction network, the cross-branch interaction module was removed, allowing the two branches to process their respective input signals independently. For each test sample to be identified, the dynamic inference process based on energy confidence is executed as follows: Acquisition: Independent output of the original branch in a two-branch network With equilibrium branch ; Confidence assessment: Real-time calculation of energy fraction for each branch and Quantify the prediction uncertainty of the current sample under each branch; Dynamic weighting: utilizing pre-trained linear mapping relationships Generate instance-level fusion weights and ; Decision output: Calculate the final probability distribution after weighted fusion. The final device prediction label is determined using the maximum a posteriori probability criterion. : In the formula, This indicates the device's predicted label.

Images