A feature memory-based incremental learning radio frequency fingerprinting method

Abstract

The present application relates to the technical field of radio frequency fingerprint identification, in particular to a kind of incremental learning radio frequency fingerprint identification method based on feature memory.For the problem that the storage overhead of historical original waveform playback is higher when new radiation source is continuously added, and the historical category recognition ability is prone to decline, the in-phase component and quadrature component of the radio frequency signal of the radiation source are collected, a noise-frequency domain composite signal representation is generated, a radio frequency fingerprint feature extraction encoder is trained and frozen, and a feature support memory bank is established according to the category;When the new radiation source is connected to the network, the historical category feature embedding vector is played back to form an incremental training batch, a comparison refinement module and an extended classifier are trained, and the radiation source recognition result is obtained by jointly constraining the classification loss, supervised comparison loss and anchor loss, which is suitable for incremental learning radio frequency fingerprint identification under dynamic device set.

Description

Technical Field

[0001] This invention relates to the field of radio frequency fingerprint recognition technology, specifically to an incremental learning radio frequency fingerprint recognition method based on feature memory. Background Technology

[0002] Radio frequency (RF) fingerprinting typically utilizes subtle signal differences arising from variations in hardware manufacturing, non-ideal characteristics of the RF front-end, and deviations in the transmission link components of wireless transmitting devices to identify radiation sources. Existing RF fingerprinting methods mostly acquire in-phase and quadrature components of the radiation source's RF signal, extract time-domain, frequency-domain, or depth features, and combine this with a classification model to determine the identity of registered radiation sources. In scenarios with newly added radiation sources, existing methods also employ incremental learning methods such as parameter regularization, historical sample playback, or category prototype maintenance to reduce the need for repeated model training.

[0003] In real-world wireless communication environments, the set of radiation source devices dynamically changes as devices join or leave the network, or as tasks change. Existing incremental learning methods based on parameter constraints often struggle to fully preserve the ability to distinguish subtle hardware differences in RF fingerprints; playback methods based on historical original waveform samples require continuously storing a large amount of historical in-phase / orthogonal RF waveform data, resulting in high storage overhead; and methods based on category prototypes are prone to a decline in historical category recognition ability due to changes in the boundaries between old and new categories as new categories are continuously added. Therefore, an incremental learning RF fingerprint recognition method based on feature memory is needed to address these issues. Summary of the Invention

[0004] To address the above problems, this invention provides the following technical solution: an incremental learning radio frequency fingerprint recognition method based on feature memory, comprising: S1. Acquire the in-phase and quadrature components of the radio frequency signal from the radiation source, perform envelope anomaly cleaning, reference envelope estimation, noise residual and frequency domain residual extraction, and generate a noise-frequency domain composite signal characterization. S2. The radio frequency fingerprint feature extraction encoder and the initial classifier are trained using the noise-frequency domain composite signal of the initially registered radiation source. After convergence, the radio frequency fingerprint feature extraction encoder is frozen to form a stable radio frequency fingerprint coordinate system. S3. Use the frozen RF fingerprint feature extraction encoder to extract the feature embedding vectors of the registered radiation source samples and store them in the feature support memory according to the category. S4. When a new radiation source is confirmed to be connected to the network, the newly added feature embedding vector is extracted using the frozen radio frequency fingerprint feature extraction encoder, and the feature embedding vectors of historical categories are replayed from the feature support memory to form an incremental training batch. S5. Train the contrast refinement module and the extended classifier with the incremental training batch, and constrain them together by the classification loss, supervised contrast loss and anchoring loss. The anchoring loss restricts the refinement deformation of historical category features. S6. The sample to be identified is processed sequentially by the frozen radio frequency fingerprint feature extraction encoder, the comparison and refinement module, and the extended classifier to obtain the radiation source identification result.

[0005] Furthermore, the characterization of the generated noise-frequency domain composite signal includes: The in-phase and quadrature components are combined into a complex baseband signal, and the amplitude envelope of the complex baseband signal is calculated. A three-standard-deviation mask is formed based on the mean and standard deviation of the amplitude envelope. Abnormal samples that exceed the mask range are cleaned to obtain the cleaned envelope. The cleaned envelope is subjected to high and low mean iterative separation to obtain an adaptive threshold, and a reference envelope is generated based on the adaptive threshold. The difference between the reference envelope and the cleaned envelope is used as the noise residual, and the difference between the spectrum of the complex baseband signal and the reference envelope is used as the frequency domain residual. The real and imaginary parts of the noise residual and the frequency domain residual are combined to form a three-channel input tensor.

[0006] Furthermore, the training of the radio frequency fingerprint feature extraction encoder and the initial classifier includes: The noise-frequency domain composite signal is input into the radio frequency fingerprint feature extraction encoder to obtain the feature embedding vector; The feature embedding vector is input into the initial classifier, and the classification cross-entropy loss is calculated based on the category labels of the initially registered radiation sources. The feature embedding vectors are normalized, and supervised contrast loss is calculated based on the cosine similarity and contrast temperature parameter between the normalized feature embedding vectors. The radio frequency fingerprint feature extraction encoder and the initial classifier are jointly trained based on the classification cross-entropy loss and the supervised contrastive loss, and the parameters of the radio frequency fingerprint feature extraction encoder are fixed to the parameters at the time of training convergence after training convergence.

[0007] Furthermore, the feature support memory includes multiple category cache units divided according to radiation source categories. Each category cache unit stores feature support records, and the feature support records include category identifiers and feature embedding vectors. Each category cache unit is set with a maximum cache quantity. When the number of feature support records of the same category exceeds the corresponding maximum cache quantity, the feature support records with higher representativeness are retained based on the similarity between feature support records of the same category, and the low representativeness feature support records with high similarity to the retained feature support records are deleted.

[0008] Furthermore, the process of forming incremental training batches includes: The newly added radiation source samples are input into the frozen radio frequency fingerprint feature extraction encoder to obtain the newly added feature embedding vector; Read the feature support records of historical categories from the feature support memory, and use the feature embedding vectors in the feature support records of historical categories as the replay feature embedding vectors; According to the batch sampling rules that include a preset replay ratio, the newly added feature embedding vector and its category label, and the replay feature embedding vector and its category label are merged to obtain the incremental training batch.

[0009] Furthermore, the comparison refinement module receives the newly added feature embedding vector and the replay feature embedding vector, and outputs refined features; the output dimension of the extended classifier corresponds to the cumulative category space after the newly added radiation source is confirmed to be in the network; the incremental training batch is only used to train the parameters of the comparison refinement module and the parameters of the extended classifier; the parameters of the frozen RF fingerprint feature extraction encoder remain unchanged during incremental training.

[0010] Furthermore, the classification loss is the classification cross-entropy loss, and the classification cross-entropy loss, supervised contrast loss, and anchoring loss constitute a combined loss function, which is expressed as follows:

[0011] in, For incremental training loss, For classification cross-entropy loss, To enable supervised comparison of losses, To anchor the loss, The weighting coefficients for the classification cross-entropy loss are... The weighting coefficients are for supervised comparison loss. This is the weighting coefficient for anchoring loss.

[0012] Furthermore, the anchoring loss uses the playback feature embedding vector in the current incremental training batch that has not been processed by the contrast refinement module as the anchor point, and uses the squared Euclidean distance between the refined feature of the playback feature embedding vector after processing by the contrast refinement module and the anchor point as the loss component, so as to limit the spatial deformation of historical category feature support in incremental training.

[0013] Furthermore, obtaining the radiation source identification result includes: Generate a noise-frequency domain composite signal representation for the sample to be identified; The corresponding noise-frequency domain composite signal is input into the frozen RF fingerprint feature extraction encoder to obtain the embedding vector of the feature to be identified. The features to be identified are embedded into the contrast refinement module after vector input training to obtain the refined features to be identified; The refined features to be identified are input into the extended classifier, and the radiation source identification result is determined based on the category corresponding to the maximum classification score output by the extended classifier. When the radiation source corresponding to the sample to be identified is confirmed as a new category and enters the next incremental stage, the feature embedding vector of the radiation source sample is extracted and written into the feature support memory, and the output dimension of the extended classifier is expanded in the next incremental stage.

[0014] Compared with the prior art, the present invention has the following beneficial effects: By performing envelope anomaly cleaning, reference envelope estimation, noise residual, and frequency domain residual extraction on the in-phase and quadrature components of the radiation source's radio frequency signal, the input representation can simultaneously include noise residual information and frequency domain residual information related to subtle hardware signal differences, and reduce the interference of anomalous samples on subsequent feature extraction. This enables the radio frequency fingerprint feature extraction encoder to learn around the signal representation that is closer to the hardware differences of the radiation source.

[0015] By training the RF fingerprint feature extraction encoder on the initial registered radiation source data and fixing its parameters after convergence, both registered and newly added radiation sources are mapped to the same stable RF fingerprint coordinate system. At the same time, by storing the feature embedding vectors of registered radiation source samples into the feature support memory according to category, the incremental training phase does not need to continuously save a large amount of historical in-phase / orthogonal RF waveform data, thereby reducing the storage resource occupation of historical sample playback and reducing the difficulty of reusing historical category features due to changes in feature coordinates.

[0016] By forming incremental training batches with newly added feature embedding vectors and replay feature embedding vectors, and training only the contrastive refinement module and the extended classifier, the learning process of new categories can simultaneously utilize the feature support of historical categories. On this basis, incremental training is constrained by classification cross-entropy loss, supervised contrastive loss, and anchoring loss, which distinguishes the feature boundaries between new categories and historical categories and limits the spatial deformation of historical category feature support during the refinement process, thereby reducing the decline in the recognition ability of historical categories when new radiation sources continue to enter the network. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method steps provided by the present invention; Figure 2 This is a schematic diagram of the initial training and incremental learning adaptation process provided by the present invention; Figure 3 This is a noise-frequency domain composite signal characterization diagram provided by the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] Please see Figures 1 to 3 , Figure 1 This is a schematic diagram of the method steps provided by the present invention; Figure 2 This is a schematic diagram of the initial training and incremental learning adaptation process provided by the present invention; Figure 3 This invention provides a noise-frequency domain composite signal characterization diagram. The invention also provides an incremental learning radio frequency fingerprinting method based on feature memory, comprising: S1. Acquire the in-phase and quadrature components of the radio frequency signal from the radiation source, perform envelope anomaly cleaning, reference envelope estimation, noise residual and frequency domain residual extraction, and generate a noise-frequency domain composite signal characterization. Specifically, the in-phase and quadrature components of the transmitted signal from the same radiation source are collected at the receiving end and aligned according to the sampling time sequence. If there are missing sampling points in the in-phase and quadrature components, the paired data corresponding to the missing sampling points are removed, ensuring that subsequent envelope calculation, frequency domain transformation, and feature extraction are based on input sequences of consistent length. After removing transient amplitude abrupt changes by envelope anomaly cleaning, the reference envelope is estimated from the cleaned envelope, and the noise residual and frequency domain residual are extracted respectively. The three-channel input tensor formed by the noise residual, the real part of the frequency domain residual, and the imaginary part of the frequency domain residual is used as the input to the subsequent RF fingerprint feature extraction encoder, enabling the subsequent model to utilize both time-domain envelope perturbation and frequency-domain difference information.

[0021] Furthermore, in one embodiment of the present invention, the generation of noise-frequency domain composite signal characterization includes: The in-phase and quadrature components are combined into a complex baseband signal, and the amplitude envelope of the complex baseband signal is calculated. A three-standard-deviation mask is formed based on the mean and standard deviation of the amplitude envelope. Abnormal samples that exceed the mask range are cleaned to obtain the cleaned envelope. The cleaned envelope is subjected to high and low mean iterative separation to obtain an adaptive threshold, and a reference envelope is generated based on the adaptive threshold. The difference between the reference envelope and the cleaned envelope is used as the noise residual, and the difference between the spectrum of the complex baseband signal and the reference envelope is used as the frequency domain residual. The real and imaginary parts of the noise residual and the frequency domain residual are combined to form a three-channel input tensor.

[0022] Specifically, record the number The in-phase components of each sampling point are Orthogonal components are The complex baseband signal is constructed as follows:

[0023] in, The sampling point number, For the first The in-phase components at each sampling point For the first Orthogonal components of each sampling point For the first The complex baseband signal corresponding to each sampling point It is the imaginary unit and satisfies The amplitude envelope of a complex baseband signal is calculated as follows:

[0024] in, For the first The amplitude envelope corresponding to each sampling point Indicates complex baseband signal The modulus is used to calculate the mean of the amplitude envelope within the same sampling segment. and standard deviation And form a mask with three times the standard deviation. When the sampling points satisfy When the sampling point is within this range, it is considered a valid sample; when the sampling point is outside this range, it is considered an abnormal sample and is cleaned. Indicates the first The absolute deviation between the amplitude envelope of each sampling point and the mean amplitude envelope This represents the outlier detection boundary corresponding to three standard deviations. During cleaning, outlier samples are removed from the envelope sequence, and the cleaned envelope is formed by interpolation of adjacent valid samples. This avoids single-point mutations affecting subsequent reference envelope estimation.

[0025] After cleaning, the envelope During the high-mean and low-mean iterative separation, the mean of the cleaned envelope is first used as the initial threshold. Samples with a value higher than or equal to the current threshold are classified as high-envelope samples, and samples with a value lower than the current threshold are classified as low-envelope samples. The mean of the high-envelope samples and the mean of the low-envelope samples are calculated separately, and the average of these two means is used to update the new threshold. This separation and update process is repeated until the change in threshold between two consecutive iterations is less than a preset iteration stop value, resulting in an adaptive threshold. This adaptive threshold is then expanded according to the sampling time sequence to form a threshold sequence consistent with the length of the cleaned envelope, and smoothed to obtain the reference envelope. .

[0026] The noise residual is determined as follows:

[0027] in, For the first The noise residual corresponding to each sampling point For the first The cleaned envelope corresponding to each sampling point For the first The reference envelope corresponding to each sampling point. For complex baseband signals. and reference envelope Perform discrete Fourier transforms on each component and calculate the spectral difference at the same frequency point to obtain the frequency domain residual. .in, Frequency point number For the first Frequency domain residuals at each frequency point Represents frequency domain residual The real part, Represents frequency domain residual The imaginary part. The noise residual. Real part of frequency domain residual and the imaginary part of the frequency domain residual After normalization to the same tensor size, they are stacked along the channel dimension to form a three-channel input tensor. Figure 3 The representations of the noise component, the real part of the noise spectrum, and the imaginary part of the noise spectrum are shown.

[0028] S2. The radio frequency fingerprint feature extraction encoder and the initial classifier are trained using the noise-frequency domain composite signal of the initially registered radiation source. After convergence, the radio frequency fingerprint feature extraction encoder is frozen to form a stable radio frequency fingerprint coordinate system. Specifically, the initially registered radiation sources are those of known categories that enter the initial training phase. S1 is performed on each registered radiation source sample to obtain the corresponding noise-frequency domain composite signal representation, which is then combined with the category label to form the initial training sample. The RF fingerprint feature extraction encoder maps the three-channel input tensor into a feature embedding vector, and the initial classifier outputs the classification score of the initially registered radiation source category based on the feature embedding vector. After training convergence, the parameters of the RF fingerprint feature extraction encoder are fixed, ensuring that samples from subsequent registered and newly added categories are mapped to the same stable RF fingerprint coordinate system, reducing the drift of historical category feature positions caused by repeated updates to the feature extraction encoder during incremental training.

[0029] Furthermore, in one embodiment of the present invention, the training of the radio frequency fingerprint feature extraction encoder and the initial classifier includes: The noise-frequency domain composite signal is input into the radio frequency fingerprint feature extraction encoder to obtain the feature embedding vector; The feature embedding vector is input into the initial classifier, and the classification cross-entropy loss is calculated based on the category labels of the initially registered radiation sources. The feature embedding vectors are normalized, and supervised contrast loss is calculated based on the cosine similarity and contrast temperature parameter between the normalized feature embedding vectors. The radio frequency fingerprint feature extraction encoder and the initial classifier are jointly trained based on the classification cross-entropy loss and the supervised contrastive loss, and the parameters of the radio frequency fingerprint feature extraction encoder are fixed to the parameters at the time of training convergence after training convergence.

[0030] Specifically, let the radio frequency fingerprint feature extraction encoder be denoted as... The initial classifier is ,in, The model parameters for the encoder used for extracting radio frequency fingerprint features. These are the model parameters for the initial classifier. The noise-frequency domain composite signal is represented. After inputting the RF fingerprint feature extraction encoder, the feature embedding vector is obtained:

[0031] in, This refers to the sample index in the training batch. For the first The noise-frequency domain composite signal representation corresponding to each sample For model parameters RF fingerprint feature extraction encoder, For the first The feature embedding vectors of each sample. The initial classifier is based on... Output the classification score for each initially registered radiation source category, and normalize it to obtain the category probability. The classification cross-entropy loss is calculated as follows:

[0032] in, For classification cross-entropy loss, This represents the number of samples in the training batch. For the first The true class label of each sample For the first Each sample in its true category The classification probability is calculated by performing L2 normalization on the feature embedding vector. ,in, Represents the feature embedding vector The 2-norm, This represents the feature embedding vector after L2 normalization. The supervised contrastive loss is determined based on positive sample pairs with the same class label and negative sample pairs with different class labels within the same batch, and is calculated as follows:

[0033] in, To enable supervised comparison of losses, This represents the number of samples in the current training batch. This serves as the sample index for comparison anchors. In the current training batch, excluding the first Apart from the sample itself, and the first A set of positive sample indexes with the same class label. For positive sample index set The number of elements in For positive sample index, In the current training batch, excluding the first Index of contrast samples other than the original sample. , , The first The sample, the first The first sample and the first The normalized feature embedding vector corresponding to each sample, with superscript... This represents the transpose of a vector. Represents the normalized feature embedding vector and The inner product between To compare temperature parameters, during training, the parameters of the RF fingerprint feature extraction encoder and the initial classifier are updated in reverse based on the weighted result of the classification cross-entropy loss and the supervised contrastive loss. When the validation set loss no longer decreases or reaches the preset number of training epochs, training is considered converged, and the parameters of the RF fingerprint feature extraction encoder are fixed to the parameters at convergence.

[0034] S3. Use the frozen RF fingerprint feature extraction encoder to extract the feature embedding vectors of the registered radiation source samples and store them in the feature support memory according to the category. Specifically, after the RF fingerprint feature extraction encoder is frozen, the initially registered radiation source samples are input again into the frozen RF fingerprint feature extraction encoder to obtain feature embedding vectors in a stable coordinate system. The feature support memory is used to store feature support vectors for historical categories, rather than continuously storing the original waveforms of historical in-phase and orthogonal components. During writing, the radiation source category is used as an index to write the feature embedding vectors of the same category to the corresponding category cache unit, enabling subsequent incremental learning adaptation stages to read historical category feature embedding vectors from the feature support memory for training.

[0035] Furthermore, in one embodiment of the present invention, the feature support memory includes multiple category cache units divided according to radiation source categories. Each category cache unit stores feature support records, and the feature support records include category identifiers and feature embedding vectors. Each category cache unit is set with a maximum cache quantity. When the number of feature support records of the same category exceeds the corresponding maximum cache quantity, the feature support records with higher representativeness are retained according to the similarity between feature support records of the same category, and the low representativeness feature support records with high similarity to the retained feature support records are deleted.

[0036] Specifically, the feature support memory uses the category identifier as the primary index, and each category cache unit stores the feature support record corresponding to that category. The feature support record requires a category identifier and a feature embedding vector as necessary fields. The category identifier is used to provide historical category labels during incremental training, and the feature embedding vector is used to replay historical category features. Each category cache unit has a maximum cache size. When the number of records supported by a certain category of features does not exceed When the number of records exceeds a certain threshold, all records are retained; when the number of records exceeds a certain threshold, all records are retained. First, the category center of the feature embedding vector for that category is calculated. Then, the similarity between each feature support record and the category center is calculated as a representativeness measure. Simultaneously, the cosine similarity between candidate feature support records and retained feature support records is calculated as a redundancy measure. When retaining feature support records, priority is given to those with high similarity to the category center and not too high similarity to already retained records. When deleting feature support records, low-representativeness records with high similarity to already retained records are deleted, thus reducing duplicate feature support when the cache size is limited.

[0037] S4. When a new radiation source is confirmed to be connected to the network, the newly added feature embedding vector is extracted using the frozen radio frequency fingerprint feature extraction encoder, and the feature embedding vectors of historical categories are replayed from the feature support memory to form an incremental training batch. Specifically, after a new radiation source is confirmed as a new category through the network access management process, S1 is performed on the new radiation source sample to obtain a noise-frequency domain composite signal representation, which is then input into the frozen RF fingerprint feature extraction encoder to obtain a new feature embedding vector. Since the RF fingerprint feature extraction encoder remains frozen, the new feature embedding vector and the feature embedding vectors of historical categories are located in the same stable RF fingerprint coordinate system. Subsequently, the feature support records of historical categories are read from the feature support memory, and the feature embedding vectors of historical categories and the new feature embedding vectors are combined to form an incremental training batch, allowing the contrast refinement module and the extended classifier to simultaneously access the feature support of historical categories while learning the new category.

[0038] Furthermore, in one embodiment of the present invention, forming incremental training batches includes: The newly added radiation source samples are input into the frozen radio frequency fingerprint feature extraction encoder to obtain the newly added feature embedding vector; Read the feature support records of historical categories from the feature support memory, and use the feature embedding vectors in the feature support records of historical categories as the replay feature embedding vectors; According to the batch sampling rules that include a preset replay ratio, the newly added feature embedding vector and its category label, and the replay feature embedding vector and its category label are merged to obtain the incremental training batch.

[0039] Specifically, after adding a new radiation source sample to the frozen RF fingerprint feature extraction encoder, a new feature embedding vector is obtained. The newly added category identifier is used as the category label for the newly added feature embedding vector. When reading feature support records of historical categories from the feature support memory, sampling is performed class by class according to the category cache unit, and the read feature embedding vector is used as the replay feature embedding vector. It reads the corresponding category identifier as the playback category label. The batch sampling rules include a preset playback ratio. This ratio is used to determine the quantity relationship between replay feature embedding vectors and newly added feature embedding vectors in an incremental training batch. If the cached quantity for a certain historical category is insufficient, the existing feature support records in the cache unit for that category are read; if there are no readable records for a certain historical category, the replay feature embedding vector for that category is not written to the current batch, and it will continue to participate in sampling after the next feature support memory update. The merged incremental training batch includes newly added feature embedding vectors, replay feature embedding vectors, and their respective category labels.

[0040] S5. Train the contrast refinement module and the extended classifier with the incremental training batch, and constrain them together by the classification loss, supervised contrast loss and anchoring loss. The anchoring loss restricts the refinement deformation of historical category features. For details, please refer to Figure 2 During the incremental learning adaptation phase, the frozen RF fingerprint feature extraction encoder no longer updates its parameters. New feature embedding vectors and replay feature embedding vectors from the incremental training batch are input to the contrast refinement module, which outputs refined features. The extended classifier receives the refined features and outputs the classification score in the cumulative class space. During training, the classification result is constrained by the classification loss, the refined features of the same class are brought closer together and the refined features of different classes are separated by supervised contrastive loss, and the spatial changes of the replay features of historical classes before and after refinement are constrained by the anchoring loss, thus creating a common constraint between learning new classes and maintaining historical classes.

[0041] Furthermore, in one embodiment provided by the present invention, the contrast refinement module receives the newly added feature embedding vector and the replay feature embedding vector, and outputs refined features; the output dimension of the extended classifier corresponds to the cumulative category space after the newly added radiation source is confirmed to be in the network; the incremental training batch is only used to train the parameters of the contrast refinement module and the parameters of the extended classifier; the parameters of the frozen radio frequency fingerprint feature extraction encoder remain unchanged during incremental training.

[0042] Specifically, the comparison and refinement module is referred to as... Expand the classifier to ,in, To compare the model parameters of the refinement module, To expand the model parameters of the classifier, the comparison and refinement module receives the newly added feature embedding vector and the replayed feature embedding vector, and outputs the refined features. ,in, The input is either the newly added feature embedding vector or the replay feature embedding vector from the contrast refinement module. To compare the refined features output by the refinement module, the output dimension of the expanded classifier is equal to the number of categories in the current cumulative category space, which includes the initially registered radiation source categories and the newly added radiation source categories that have been confirmed to be included in the network. During incremental training, backpropagation only updates the parameters of the comparison refinement module. and extended classifier parameters Do not update the encoder parameters for RF fingerprint feature extraction Therefore, the base coordinates of the feature embedding vector do not change with training on new categories, while the discriminative boundary between new and historical categories is handled by the contrastive refinement module and the extended classifier.

[0043] Furthermore, in one embodiment provided by the present invention, the classification loss is the classification cross-entropy loss, and the classification cross-entropy loss, supervised contrast loss, and anchoring loss constitute a combined loss function, which is expressed as follows:

[0044] in, For incremental training loss, For classification cross-entropy loss, To enable supervised comparison of losses, To anchor the loss, These are the weighting coefficients for the classification cross-entropy loss. The weighting coefficients are for supervised comparison loss. This is the weighting coefficient for anchoring loss.

[0045] Specifically, the classification cross-entropy loss is calculated based on the cumulative class space classification probability output by the expanded classifier and the class labels in the incremental training batch; the supervised contrastive loss is calculated based on the cosine similarity between refined features, class labels, and contrast temperature parameters; and the anchoring loss is calculated based on the distance between the replay feature embedding vectors before and after processing by the contrastive refinement module. These three loss terms are combined according to their weights to form the incremental training loss, which is used to update the contrastive refinement module and the expanded classifier. In this embodiment, the weights are set based on the changes in classification loss on the training and validation sets and the stability of historical class recognition, and remain consistent within the same incremental training phase to avoid frequent changes in the objective function during training.

[0046] Furthermore, in one embodiment provided by the present invention, the anchoring loss uses the playback feature embedding vector in the current incremental training batch that has not been processed by the contrast refinement module as the anchor point, and uses the squared Euclidean distance between the refined feature of the playback feature embedding vector after being processed by the contrast refinement module and the anchor point as the loss component, so as to limit the spatial deformation of historical category feature support in incremental training.

[0047] Specifically, the sample indices corresponding to the replay feature embedding vectors in the current incremental training batch are used to form a replay sample index set. For any replay sample index The feature embedding vector of the replay sample before the comparison refinement module is denoted as... And using its vector position as the anchor point, the refined feature of the replay feature embedding vector after processing by the comparison and refinement module is denoted as... Anchoring loss is calculated as follows:

[0048] in, To anchor the loss, This is the set of replay sample indices in the current incremental training batch. For the replay sample index set The number of elements in the dataset, i.e., the number of replay feature embedding vectors in the current incremental training batch. For replay sample index, For the first The feature embedding vectors of the replay samples before the comparison refinement module. For the first Refined features of each replay sample after comparison and refinement module processing Indicates refinement features With corresponding anchor points The squared L2 norm between them is the squared Euclidean distance between them. This loss is calculated only for the replay feature embedding vector, ensuring that the historical category feature support remains close to its original position formed by the frozen encoder after the contrastive refinement module; when the historical category refined features deviate from the anchor point, the squared Euclidean distance increases and improves the anchoring loss, thereby limiting the spatial deformation of the historical category feature support in the reverse update.

[0049] S6. The sample to be identified is processed sequentially by the frozen radio frequency fingerprint feature extraction encoder, the comparison and refinement module, and the extended classifier to obtain the radiation source identification result.

[0050] Specifically, when the sample to be identified enters the identification stage, a noise-frequency domain composite signal representation is first generated according to S1, and then input into the frozen RF fingerprint feature extraction encoder to obtain the feature embedding vector to be identified. Since the parameters of the RF fingerprint feature extraction encoder remain fixed, the feature embedding vector to be identified and the historical category and newly added category feature embedding vectors formed during the training stage are located in the same stable RF fingerprint coordinate system. Subsequently, the feature embedding vector to be identified is input into the trained contrast refinement module to obtain the refined features to be identified, and the extended classifier outputs the classification scores of each category in the cumulative category space. The radiation source identification result is determined based on the classification scores.

[0051] Furthermore, in one embodiment of the present invention, obtaining the radiation source identification result includes: Generate a noise-frequency domain composite signal representation for the sample to be identified; The corresponding noise-frequency domain composite signal is input into the frozen RF fingerprint feature extraction encoder to obtain the embedding vector of the feature to be identified. The features to be identified are embedded into the contrast refinement module after vector input training to obtain the refined features to be identified; The refined features to be identified are input into the extended classifier, and the radiation source identification result is determined based on the category corresponding to the maximum classification score output by the extended classifier. When the radiation source corresponding to the sample to be identified is confirmed as a new category and enters the next incremental stage, the feature embedding vector of the radiation source sample is extracted and written into the feature support memory, and the output dimension of the extended classifier is expanded in the next incremental stage.

[0052] Specifically, the same signal representation generation process as the training phase is performed on the sample to be identified to obtain the corresponding noise-frequency domain composite signal representation; this representation is then input into the frozen RF fingerprint feature extraction encoder to obtain the feature embedding vector to be identified. ;Will Input the trained comparison refinement module to obtain the refined features to be identified. ;Will The extended classifier is input to obtain classification scores for each category. The category corresponding to the maximum value in the classification score is used as the radiation source identification result. When the sample to be identified is confirmed as a new category through the external network access confirmation process and enters the next incremental stage, the frozen RF fingerprint feature extraction encoder is used to extract the feature embedding vector of the radiation source sample, and the feature embedding vector, along with the new category identifier, is written into the corresponding category cache unit of the feature support memory. At the same time, the new category is included in the cumulative category space of the next incremental stage, so that the output dimension of the extended classifier in the next incremental stage increases by the corresponding category channel.

[0053] In one embodiment of the continuous operation, an initial training set is first established using initially registered radiation source samples, and a noise-frequency domain composite signal representation is generated for each sample. After the training of the RF fingerprint feature extraction encoder and the initial classifier converges, the parameters of the RF fingerprint feature extraction encoder are frozen, and the feature embedding vectors of the initially registered radiation source samples are written into the feature support memory according to their categories. When a new radiation source is confirmed to be added to the network, the new sample obtains a new feature embedding vector through the frozen RF fingerprint feature extraction encoder, and the feature support memory replays the feature embedding vectors of historical categories. The two are combined into an incremental training batch according to a batch sampling rule that includes a preset replay ratio. After this batch is input into the contrast refinement module and the extended classifier, it is trained by classification cross-entropy loss, supervised contrast loss, and anchoring loss. Among them, the anchoring loss uses the replay feature embedding vector as the anchor point to constrain the offset of the historical category refinement features. After the current incremental stage training is completed, the sample to be identified sequentially passes through the frozen RF fingerprint feature extraction encoder, the contrast refinement module, and the extended classifier, and the radiation source identification result is formed by the category corresponding to the maximum classification score.

[0054] In one embodiment of the logic verification scenario, historical category samples from the same batch only participate in incremental training using feature support records, without reading the original in-phase and quadrature RF waveforms. New category samples are entered into the same stable RF fingerprint coordinate system through the frozen RF fingerprint feature extraction encoder. If the historical category playback features deviate from their anchor points after processing by the contrast refinement module, the anchoring loss increases with the square of the Euclidean distance and participates in the inverse constraint of the incremental training loss. If the distance between the new category and the historical category in the refined feature space is insufficient, the supervised contrast loss participates in separating features of different categories and tightening features of the same category. This logic verification is used to illustrate the execution relationship between feature support playback, contrast refinement, and anchoring constraints, and is not intended to limit actual test data or engineering operation data.

[0055] 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.

[0056] 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.

Claims

1. An incremental learning radio frequency fingerprint recognition method based on feature memory, characterized in that, include: S1. Acquire the in-phase and quadrature components of the radio frequency signal from the radiation source, perform envelope anomaly cleaning, reference envelope estimation, noise residual and frequency domain residual extraction, and generate a noise-frequency domain composite signal characterization. S2. The radio frequency fingerprint feature extraction encoder and the initial classifier are trained using the noise-frequency domain composite signal of the initially registered radiation source. After convergence, the radio frequency fingerprint feature extraction encoder is frozen to form a stable radio frequency fingerprint coordinate system. S3. Use the frozen RF fingerprint feature extraction encoder to extract the feature embedding vectors of the registered radiation source samples and store them in the feature support memory according to the category. S4. When a new radiation source is confirmed to be connected to the network, the newly added feature embedding vector is extracted using the frozen radio frequency fingerprint feature extraction encoder, and the feature embedding vectors of historical categories are replayed from the feature support memory to form an incremental training batch. S5. Train the contrast refinement module and the extended classifier with the incremental training batch, and constrain them together by the classification loss, supervised contrast loss and anchoring loss. The anchoring loss restricts the refinement deformation of historical category features. S6. The sample to be identified is processed sequentially by the frozen radio frequency fingerprint feature extraction encoder, the comparison and refinement module, and the extended classifier to obtain the radiation source identification result.

2. The incremental learning radio frequency fingerprint recognition method based on feature memory according to claim 1, characterized in that, The generated noise-frequency domain composite signal characterization includes: The in-phase and quadrature components are combined into a complex baseband signal, and the amplitude envelope of the complex baseband signal is calculated. A three-standard-deviation mask is formed based on the mean and standard deviation of the amplitude envelope. Abnormal samples that exceed the mask range are cleaned to obtain the cleaned envelope. The cleaned envelope is subjected to high and low mean iterative separation to obtain an adaptive threshold, and a reference envelope is generated based on the adaptive threshold. The difference between the reference envelope and the cleaned envelope is used as the noise residual, and the difference between the spectrum of the complex baseband signal and the reference envelope is used as the frequency domain residual. The real and imaginary parts of the noise residual and the frequency domain residual are combined to form a three-channel input tensor.

3. The incremental learning radio frequency fingerprint recognition method based on feature memory according to claim 1, characterized in that, The training radio frequency fingerprint feature extraction encoder and initial classifier include: The noise-frequency domain composite signal is input into the radio frequency fingerprint feature extraction encoder to obtain the feature embedding vector; The feature embedding vector is input into the initial classifier, and the classification cross-entropy loss is calculated based on the category labels of the initially registered radiation sources. The feature embedding vectors are normalized, and supervised contrast loss is calculated based on the cosine similarity and contrast temperature parameter between the normalized feature embedding vectors. The radio frequency fingerprint feature extraction encoder and the initial classifier are jointly trained based on the classification cross-entropy loss and the supervised contrastive loss, and the parameters of the radio frequency fingerprint feature extraction encoder are fixed to the parameters at the time of training convergence after training convergence.

4. The incremental learning radio frequency fingerprint recognition method based on feature memory according to claim 1, characterized in that, The feature support memory includes multiple category cache units divided according to radiation source categories. Each category cache unit stores feature support records, which include category identifiers and feature embedding vectors. Each category cache unit is set with a maximum cache quantity. When the number of feature support records of the same category exceeds the corresponding maximum cache quantity, the feature support records with higher representativeness are retained based on the similarity between feature support records of the same category, and the low representativeness feature support records with high similarity to the retained feature support records are deleted.

5. The incremental learning radio frequency fingerprint recognition method based on feature memory according to claim 4, characterized in that, The process of forming incremental training batches includes: The newly added radiation source samples are input into the frozen radio frequency fingerprint feature extraction encoder to obtain the newly added feature embedding vector; Read the feature support records of historical categories from the feature support memory, and use the feature embedding vectors in the feature support records of historical categories as the replay feature embedding vectors; According to the batch sampling rules that include a preset replay ratio, the newly added feature embedding vector and its category label, and the replay feature embedding vector and its category label are merged to obtain the incremental training batch.

6. The incremental learning radio frequency fingerprint recognition method based on feature memory according to claim 5, characterized in that, The comparison refinement module receives the newly added feature embedding vector and the replay feature embedding vector, and outputs refined features; The output dimension of the extended classifier corresponds to the cumulative category space after the newly added radiation source is confirmed to be included in the network. The incremental training batch is only used to train the parameters of the comparison and refinement module and the parameters of the extended classifier. The parameters of the frozen RF fingerprint feature extraction encoder remain unchanged during the incremental training.

7. The incremental learning radio frequency fingerprint recognition method based on feature memory according to claim 1, characterized in that, The classification loss is the classification cross-entropy loss. The classification cross-entropy loss, supervised contrastive loss, and anchoring loss constitute a combined loss function, which is expressed as follows: ;in, For incremental training loss, For classification cross-entropy loss, To enable supervised comparison of losses, To anchor the loss, These are the weighting coefficients for the classification cross-entropy loss. The weighting coefficients are for supervised comparison loss. This is the weighting coefficient for anchoring loss.

8. The incremental learning radio frequency fingerprint recognition method based on feature memory according to claim 7, characterized in that, The anchoring loss uses the playback feature embedding vector in the current incremental training batch that has not been processed by the contrast refinement module as the anchor point, and uses the squared Euclidean distance between the refined feature of the playback feature embedding vector after processing by the contrast refinement module and the anchor point as the loss component, so as to limit the spatial deformation of historical category feature support in incremental training.

9. The incremental learning radio frequency fingerprint recognition method based on feature memory according to claim 1, characterized in that, The obtained radiation source identification results include: Generate a noise-frequency domain composite signal representation for the sample to be identified; The corresponding noise-frequency domain composite signal is input into the frozen RF fingerprint feature extraction encoder to obtain the embedding vector of the feature to be identified. The features to be identified are embedded into the contrast refinement module after vector input training to obtain the refined features to be identified; The refined features to be identified are input into the extended classifier, and the radiation source identification result is determined based on the category corresponding to the maximum classification score output by the extended classifier. When the radiation source corresponding to the sample to be identified is confirmed as a new category and enters the next incremental stage, the feature embedding vector of the radiation source sample is extracted and written into the feature support memory, and the output dimension of the extended classifier is expanded in the next incremental stage.

Images