A semi-supervised radio frequency fingerprint deep clustering identification method based on adaptive loss

Abstract

The application discloses a kind of semi-supervised radio frequency fingerprint depth clustering identification method based on adaptive loss, first, radio frequency IQ data is collected and preprocessed, radio frequency training sample data is obtained and special data set is constructed, then depth learning feature extraction network and adaptive loss function are constructed, through model training stage, adaptive weight coefficient is determined, and using fixed weight is retrained to network, the convergent model weight file is obtained, the data is inferred using the trained model, the spatial feature vector is extracted, and the feature vectors of different IQ data are calculated for similarity, relationship matrix is constructed, finally, it is handled by heap, the automatic aggregation of the same kind of radiation source is realized by connected component extraction, and the final radio frequency target recognition result is output.The method of the application is suitable for in the actual deployment environment of labeled sample scarcity and unknown transmitting device, realizes the high-precision identification of known radiation source and the automatic discovery and clustering of unknown radiation source.

Description

Technical Field

[0001] This invention belongs to the field of radio frequency fingerprint recognition technology, specifically relating to a semi-supervised radio frequency fingerprint deep clustering recognition method based on adaptive loss. Background Technology

[0002] Radio frequency fingerprinting (RFF) technology utilizes unique signal features introduced by hardware defects in transmitting equipment to authenticate devices, serving as a crucial means of physical layer security. In recent years, deep learning has made significant progress in this field, with convolutional neural network-based methods achieving extremely high recognition accuracy under closed-set, fully supervised conditions. However, in real-world wireless environments, acquiring large amounts of high-quality labeled signals is costly, especially in non-cooperative scenarios where complete labeling is difficult to achieve, leading to a sharp decline in the performance of supervised learning methods as the labeling ratio decreases.

[0003] In practical deployment, even more complex challenges arise: new transmitters can appear at any time, and existing closed-set classifiers may forcibly classify unknown targets into known categories, leading to misjudgments; while existing open-set recognition methods can reject unknown samples, they often classify all unknowns into a single "unknown" category, failing to distinguish between multiple different unknown transmitters. Simultaneously, factors such as equipment aging, changes in ambient temperature, and alterations in the electromagnetic environment can cause signal feature drift, degrading the performance of pre-trained models. Furthermore, the distribution of data collected at different times varies, making it difficult for fixed loss weights to adapt to the dynamic training process, further exacerbating the recognition difficulty.

[0004] Existing semi-supervised learning methods (such as FixMatch and DCR) improve classification performance using a small amount of labeled data and a large amount of unlabeled data, but they typically assume that the unlabeled data only contains known categories, making them unable to handle the discovery of unknown categories. While contrastive learning methods can learn discriminative features, they lack clustering mechanisms for unknown categories. More importantly, most existing methods use fixed loss weights, failing to dynamically adjust the balance between supervised and unsupervised information based on the training state. This makes them susceptible to noise interference in the early stages of training and unable to fully exploit the data structure in the later stages.

[0005] Therefore, there is an urgent need for a radio frequency fingerprinting method that can simultaneously identify known categories and automatically discover multiple unknown categories in semi-supervised scenarios, and has adaptive loss adjustment capabilities. Summary of the Invention

[0006] To address the aforementioned technical issues, this invention proposes a semi-supervised radio frequency fingerprint deep clustering identification method based on adaptive loss. This method is suitable for practical deployment environments where labeled samples are scarce and unknown transmitting devices exist, enabling high-precision identification of known radiation sources and automatic discovery and clustering of unknown radiation sources.

[0007] The technical solution adopted in this invention is: a semi-supervised radio frequency fingerprint deep clustering recognition method based on adaptive loss, the specific steps of which are as follows:

[0008] S1. Obtain radio frequency signal IQ data at different time periods, perform preprocessing to obtain radio frequency training sample data, and construct a dedicated dataset.

[0009] S2. Based on step S1, construct a deep learning feature extraction network and a deep learning adaptive loss function;

[0010] The input to the deep learning feature extraction network is the IQ data of the radio frequency signal, and the input size is 1× The network uses blocks as the basic feature extraction modules. Each block includes a convolutional layer, a batch normalization layer, and a ReLU activation function. After feature extraction through multiple blocks, the network is finally connected to a layer containing... A fully connected layer of neurons uses a softmax function to make each dimension of the vector represent a different spatial mapping location.

[0011] in, This indicates the preset feature space dimension.

[0012] S3. Based on the deep learning feature extraction network constructed in step S2, input the special dataset constructed in step S1 to train the model and determine the model loss weights.

[0013] S4. Based on the loss weights determined in step S3, retrain the deep learning feature extraction network using a deep learning adaptive loss function with fixed weights until the loss function converges, and save the final trained model weight file.

[0014] Among them, the final total loss function The expression is as follows:

[0015] ;

[0016] in, This represents the supervised cross-entropy loss. Represents the instance cross-entropy loss. Indicates the contrast learning loss. and This represents the final value of the adaptively adjusted weighting coefficients.

[0017] S5. Based on step S4, use a deep learning feature extraction network and a trained model weight file to perform spatial feature inference and prediction on all data to obtain the feature vector of each sample. Then, sum the feature vectors of the samples corresponding to each IQ data and calculate the average to obtain the overall representation feature vector of the IQ data.

[0018] S6. Based on step S5, perform similarity distance calculation on the feature vectors of different IQ data to obtain the spatial relationship matrix;

[0019] Given N IQ data points, we calculate the similarity distance between the feature vectors of these N IQ data points, resulting in an N*N comparison distance matrix. The ( )th element of the matrix... The value of ) is the first The IQ data and the first Spatial distance between IQ data points.

[0020] Then set the threshold as If the spatial distance is greater than or equal to If the result is 1, then set it to 1; otherwise, set it to 0 to obtain the spatial relation matrix.

[0021] S7. Based on step S6, the spatial relationship matrix is ​​split into clusters, and the automatic aggregation of similar radiation sources is achieved by extracting connected components to obtain the radio frequency fingerprint recognition results of different IQ data.

[0022] The clustering process involves determining whether two IQ data are related to each other or can be related through other IQ files. If so, they are determined to be the same radiation source, and the same radiation sources are grouped together.

[0023] Furthermore, step S1 is specifically as follows:

[0024] S11. Acquire radio frequency signal IQ data at different time periods;

[0025] Each time period contains a different number of transmitters, with some transmitters appearing only in certain time periods. All signals constitute the training dataset, with a small number of samples bearing known class labels, denoted as the annotation set. The remaining samples are unlabeled and are referred to as the unlabeled set. Furthermore, the set does not contain unknown transmitters. Assume the total number of known categories is... .

[0026] S12. After preprocessing the data obtained in step S11, radio frequency training sample data is obtained.

[0027] Each IQ signal is normalized to zero mean to eliminate amplitude differences while preserving hardware fingerprint characteristics. The normalization formula is as follows:

[0028] ;

[0029] in, This represents the signal before normalization. This represents the normalized signal, where μ(x) and σ(x) represent the mean and standard deviation of the signal, respectively. No carrier frequency offset correction, phase alignment, or IQ imbalance compensation is performed during the processing.

[0030] Then based on the known number of categories And to estimate the number of possible unknown categories, and to set the dimension of the feature space. Its satisfaction Furthermore, this dimension will be used as the output dimension of the last fully connected layer of the deep learning model, and each dimension corresponds to the probability of a cluster center.

[0031] in, This indicates the maximum number of unknown categories preset.

[0032] Each normalized signal is then divided into continuous slices of fixed length, with a slice length of... Each slice is considered an independent sample. Different slices of the same signal are considered positive sample pairs. The normalized slice samples are the radio frequency training sample data.

[0033] S13. Construct a dedicated dataset based on step S12;

[0034] Define the complete tag space Establish mapping This ensures that each unlabeled signal has a globally unique pseudo-label. Then, the radio frequency training sample data and their corresponding labels are combined to form a dedicated dataset.

[0035] in, Represents a known set of class labels. This represents a unique individual pseudo-label assigned to each unlabeled signal.

[0036] Furthermore, step S2 is specifically as follows:

[0037] First, a deep learning feature extraction network is automatically constructed based on the label space of the dedicated dataset. Then, a deep learning adaptive loss function is constructed, and the total loss function is... The weighted sum of the three losses is expressed as follows:

[0038] ;

[0039] in, This represents the supervised cross-entropy loss. Represents the instance cross-entropy loss. Indicates the contrast learning loss. and This represents the adaptively adjusted weight coefficients, with an initial value set to... .

[0040] For the annotation set For the samples in the dataset, the supervised cross-entropy loss is calculated using their true known class labels, as shown in the following expression:

[0041] ;

[0042] in, This represents the set of indices for the labeled samples in the current batch. This represents the true label of the sample. This indicates the probability that the model predicts the sample belongs to the true class. This indicates the number of labeled signals selected in the current batch. This indicates the index of the labeled signal.

[0043] For unlabeled sets For the samples in step S1, the instance cross-entropy loss is calculated using the unique individual pseudo-labels assigned in step S1, as shown in the following expression:

[0044] ;

[0045] in, This represents the set of indices for unlabeled samples in the current batch. This represents the unique pseudo-label corresponding to the sample. This indicates the number of unlabeled signals selected in the current batch. This indicates that the signal index is not labeled.

[0046] The expression for calculating the contrastive learning loss is as follows:

[0047] ;

[0048] in, This represents the distance loss for positive sample pairs, which forces the Euclidean distance between positive sample pairs to decrease; This represents the distance loss for negative sample pairs, which forces the Euclidean distance between negative sample pairs to increase. This represents the similarity loss of positive sample pairs, forcing the cosine similarity of positive sample pairs to increase; This represents the similarity loss of negative sample pairs, forcing the cosine similarity of negative sample pairs to decrease. The specific expressions are as follows:

[0049] ;

[0050] ;

[0051] ;

[0052] ;

[0053] in, This represents a small constant to prevent overflow in logarithmic operations. The L1 norm of a matrix is ​​represented by the Euclidean distance metric. The value ranges from 0 to 1; the cosine similarity metric is... The value ranges from 0 to 1. This indicates the transpose operation. This represents a mask matrix for positive sample pairs, with dimensions N*N, where N represents the dimension of the mask matrix. With sample When from the same signal Otherwise, it is 0. Describes the mask matrix of negative sample pairs, satisfying Where 1 is an all-one matrix, and I represents the identity matrix, i.e., when And when the two samples come from different signals , and and represent the L1 norm of the positive and negative mask matrices, respectively, which is the sum of all elements in the matrix.

[0054] Furthermore, step S3 is specifically as follows:

[0055] First, set the threshold for the known class. Unknown class threshold and update factor and The deep learning feature extraction network and loss function are trained on a dedicated dataset for one cycle to count the number of classes with consistent predictions in known class blocks. The number of different pseudo-labels activated in unlabeled blocks And reset based on the results and After initializing the model weights, retrain until the detection is successful. The trend of change reverses, that is, after a continuous increase, it decreases for the first time, or after a continuous decrease, it increases for the first time, and finally determines... and The value is .

[0056] in, Indicates the weight reduction factor. This represents the weight amplification factor; for the supervised loss weight α, if This indicates that the known classes are stable, and the weights of the supervised loss should be appropriately reduced to prevent overfitting. Let Otherwise, increase the weight of the supervision loss, making For the instance loss weight β, if This indicates that the unlabeled samples are sufficiently dispersed, and the instance loss weight should be appropriately reduced. Otherwise, increase the instance loss weight, making .

[0057] The beneficial effects of this invention are as follows: The method of this invention first collects and preprocesses radio frequency (RF) IQ data to obtain RF training sample data and constructs a dedicated dataset. Then, it constructs a deep learning feature extraction network and an adaptive loss function. During the model training phase, it dynamically monitors the prediction consistency of known classes and the dispersion of unlabeled classes, adaptively determines the weight coefficients of supervised loss and instance loss, and then retrains the network using fixed weights to obtain a converged model weight file. The trained model is then used to infer the data, extract spatial feature vectors, and calculate the similarity of feature vectors from different IQ data to construct a relation matrix. Finally, the relation matrix is ​​processed by stacking, and automatic aggregation of similar radiation sources is achieved through connected component extraction, outputting the final RF target recognition result. The method of this invention not only has high accuracy in radio frequency fingerprint closed-set recognition, but also has a certain recognition effect on unknown radio frequency targets. At the same time, it has a certain effect on solving feature drift to a certain extent. It realizes full automation from raw data to recognition results. Especially in the scenario where labeled data is scarce and there are unknown transmitters, it can simultaneously complete the recognition of known classes and the discovery of unknown classes. It is suitable for actual deployment environments where labeled samples are scarce and there are unknown transmitting devices, so as to achieve high-precision recognition of known radiation sources and automatic discovery and clustering of unknown radiation sources.

[0058] The method of this invention dynamically adjusts the weight coefficients of supervised loss and instance loss by periodically monitoring the consistency of predictions for known classes and the dispersion of unlabeled classes. When the predictions for known classes are stable, the supervised weight is appropriately reduced to prevent overfitting; when unstable, the supervised guidance is enhanced. When the dispersion of unlabeled data is insufficient, the instance loss is enhanced to promote separation; when excessively dispersed, it is appropriately reduced to avoid interference. This mechanism achieves a dynamic balance between supervised and unsupervised information, significantly improving training stability and final recognition performance. The method of this invention assigns a unique pseudo-label to each unlabeled signal, and forces different signals to remain dispersed in the feature space through instance loss, preventing collapse to known categories. In the later stages of training, stability-driven consensus clustering is used to collect prediction results from multiple stable rounds to construct a co-occurrence matrix, and automatically aggregates multiple unknown categories through graph connectivity component extraction. This method does not require a preset number of categories and can adaptively discover any number of novel emitters, breaking through the limitation of traditional methods that classify unknowns into a single category. The method of this invention simultaneously introduces Euclidean distance and cosine similarity comparison constraints in the probability space, bringing positive sample pairs closer while pushing negative sample pairs further apart, ensuring that samples of the same type are tightly clustered and samples of different types are fully separated. By eliminating early unstable prediction interference through stable rounds of screening, and combining a multi-level evaluation system, a three-tiered recognition reliability system (strong, medium, and weak) is provided. This method exhibits good robustness to feature drift caused by equipment aging and environmental changes, and can maintain stable recognition performance even in dynamic and non-stationary environments. Attached Figure Description

[0059] Figure 1 This is a flowchart of a semi-supervised radio frequency fingerprint deep clustering recognition method based on adaptive loss according to the present invention.

[0060] Figure 2 This is a schematic diagram of the deep learning feature extraction network structure in an embodiment of the present invention.

[0061] Figure 3 This is a schematic diagram of the adaptive loss function structure in an embodiment of the present invention.

[0062] Figure 4 This is a diagram showing the feature clustering effect based on semi-supervised recognition in an embodiment of the present invention. Detailed Implementation

[0063] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0064] like Figure 1 The flowchart of a semi-supervised radio frequency fingerprint deep clustering recognition method based on adaptive loss is shown below. The specific steps are as follows:

[0065] S1. Obtain radio frequency signal IQ data at different time periods, perform preprocessing to obtain radio frequency training sample data, and construct a dedicated dataset.

[0066] S11. Acquire radio frequency signal IQ data at different time periods;

[0067] Each time period contains a different number of transmitters, with some transmitters appearing only in certain time periods. All signals constitute the training dataset, with a small number of samples bearing known class labels, denoted as the annotation set. The remaining samples are unlabeled and are referred to as the unlabeled set. Furthermore, the set does not contain unknown transmitters. Assume the total number of known categories is... .

[0068] In this example, a publicly available LoRa RF fingerprint dataset is used. This dataset contains 25 Pycom SX1276 transmitters, with an Ettus USRP B210 receiver, a center frequency of 915 MHz, a sampling rate of 1 MS / s, and a complex IQ signal format. An outdoor subset of data from 10 transmitters is selected for the experiment, with 5 representing known classes and 5 representing unknown classes.

[0069] In this embodiment, IQ signals from 10 radio frequency transmitters are acquired, and 10 IQ data points are collected from each transmitter. 50% of the samples are randomly selected and assigned known class labels (corresponding to 5 known transmitters), while the remaining 50% of the samples are treated as unlabeled data (including samples from 5 known transmitters and 5 unknown transmitters).

[0070] S12. After preprocessing the data obtained in step S11, radio frequency training sample data is obtained.

[0071] Each IQ signal is normalized to zero mean to eliminate amplitude differences while preserving hardware fingerprint characteristics. The normalization formula is as follows:

[0072] ;

[0073] in, This represents the signal before normalization. The normalized signal is represented by μ(x) and σ(x), which represent the mean and standard deviation of the signal, respectively. Furthermore, no carrier frequency offset correction, phase alignment, or IQ imbalance compensation is performed during processing to avoid damaging the device's unique hardware fingerprint information.

[0074] Then based on the known number of categories =5 and the estimated number of unknown categories are used to define the feature space dimension. =80, which satisfies Furthermore, this dimension will be used as the output dimension of the last fully connected layer of the deep learning model, and each dimension corresponds to the probability of a cluster center.

[0075] in, This indicates the maximum number of unknown categories preset; in this embodiment, it is 75.

[0076] Each normalized signal is then divided into continuous slices of fixed length, with a slice length of... Each signal generates 2000 slices. Each slice is considered an independent sample. Different slices of the same signal are considered positive sample pairs. This operation not only increases the number of samples but also provides natural positive sample pairs for subsequent contrastive learning. The normalized slice samples are the radio frequency training sample data.

[0077] S13. Construct a dedicated dataset based on step S12;

[0078] First, define the complete tag space. In this embodiment, the known class labels are 1 to 5, and there are a total of 75 signals in the unlabeled set. A unique individual pseudo-label 6 to 80 is assigned to each of these signals. Then, a mapping is established. Each signal ID is recorded with a pseudo-label, ensuring that each unlabeled signal has a globally unique pseudo-label. This design treats each unlabeled signal as an independent category in the initial stage, preventing it from collapsing into known categories. Finally, the radio frequency training sample data and their corresponding labels are combined to form a dedicated dataset.

[0079] in, Represents a known set of class labels. This represents a unique individual pseudo-label assigned to each unlabeled signal.

[0080] S2. Based on step S1, construct a deep learning feature extraction network and a deep learning adaptive loss function;

[0081] The deep learning feature extraction network structure designed in this embodiment is as follows: Figure 2 As shown, the network input is the IQ data of the radio frequency signal, with an input size of 1×8192. The network uses blocks as the basic feature extraction module, and each block contains a convolutional layer, a batch normalization layer, and a ReLU activation function. The specific structure is as follows: First, it passes through 5 consecutive block modules, each using a 1×3 convolutional kernel and 512 output channels; then it passes through a block module, also using a 1×3 convolutional kernel and 512 channels, but followed by a 1×2 max pooling layer and a dropout layer. This structure consisting of 5 ordinary blocks and 1 block with pooling and dropout is repeated 7 times. Then, it passes through 5 consecutive ordinary block2 modules again, and finally through a block module with 1×2 pooling and dropout. After all convolutional operations are completed, the resulting feature map is flattened and connected to a fully connected layer containing 512 neurons with the ReLU activation function; finally, it is connected to a fully connected layer containing M neurons with the softmax activation function, outputting the feature map of each sample. The probability distribution of each potential category.

[0082] in, This indicates the preset feature space dimension.

[0083] Then, a deep learning adaptive loss function is constructed. In this embodiment, the loss function is as follows: Figure 3 As shown, the total loss function consists of three parts: supervised cross-entropy loss, instance cross-entropy loss, and contrastive alignment loss. The contrastive alignment loss is further decomposed into four sub-items: positive sample pair distance loss, negative sample pair distance loss, positive sample pair similarity loss, and negative sample pair similarity loss. Supervised cross-entropy loss applies to labeled samples, ensuring that known categories form clear cluster centers. Instance cross-entropy loss assigns a unique pseudo-label to each unlabeled signal, forcing different signals to remain dispersed in the feature space. Contrastive alignment loss provides refined constraints on sample pair relationships in the probability space, controlling the Euclidean distance between samples through distance loss and the cosine similarity between samples through similarity loss. The synergistic effect of these four factors ensures that similar samples cluster tightly and dissimilar samples are fully separated.

[0084] Then the total loss function The expression is as follows:

[0085] ;

[0086] in, This represents the supervised cross-entropy loss. Represents the instance cross-entropy loss. Indicates the contrast learning loss. and This represents the adaptively adjusted weight coefficients, with an initial value set to... .

[0087] For the annotation set For the samples in the dataset, the supervised cross-entropy loss is calculated using their true known class labels to anchor the decision boundary of the known classes, as shown in the following expression:

[0088] ;

[0089] in, This represents the set of indices for the labeled samples in the current batch. This represents the true label of the sample. This indicates the probability that the model predicts the sample belongs to the true class. This indicates the number of labeled signals selected in the current batch. This represents the labeled signal index. This loss ensures that known categories have clear cluster centers in the feature space.

[0090] For unlabeled sets For the samples in step S1, the instance cross-entropy loss is calculated using the unique individual pseudo-label assigned in step S1. This forces different unlabeled signals to disperse in the feature space, preventing collapse to a known category. The expression is as follows:

[0091] ;

[0092] in, This represents the set of indices for unlabeled samples in the current batch. This represents the unique pseudo-label corresponding to the sample. This indicates the number of unlabeled signals selected in the current batch. This represents the index of the unlabeled signal. This loss ensures that each unlabeled signal has an independent pseudo-cluster center in the initial stage.

[0093] The expression for calculating the contrastive learning loss is as follows:

[0094] ;

[0095] in, This represents the distance loss for positive sample pairs, which forces the Euclidean distance between positive sample pairs to decrease; This represents the distance loss for negative sample pairs, which forces the Euclidean distance between negative sample pairs to increase. This represents the similarity loss of positive sample pairs, forcing the cosine similarity of positive sample pairs to increase; This represents the similarity loss of negative sample pairs, forcing the cosine similarity of negative sample pairs to decrease. The specific expressions are as follows:

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] in, This represents a small constant to prevent overflow in logarithmic operations. The L1 norm of a matrix is ​​represented by the Euclidean distance metric. The value ranges from 0 to 1, with larger values ​​indicating greater differences between the two probability vectors; the cosine similarity metric is... The value ranges from 0 to 1; a larger value indicates that the two probability vectors are more similar. This indicates the transpose operation. This represents a mask matrix for positive sample pairs, with dimensions N*N, where N represents the dimension of the mask matrix. With sample When from the same signal Otherwise, it is 0. Describes the mask matrix of negative sample pairs, satisfying Where 1 is an all-one matrix, and I represents the identity matrix, i.e., when And when the two samples come from different signals , and and represent the L1 norm of the positive and negative mask matrices, which is the sum of all elements in the matrix. It is used to normalize the loss and balance the difference in the number of positive and negative sample pairs.

[0101] S3. Based on the deep learning feature extraction network constructed in step S2, input the special dataset constructed in step S1 to train the model and determine the model loss weights.

[0102] First, set the threshold for the known class. Unknown class threshold and update factor and The deep learning feature extraction network and loss function are trained on a dedicated dataset for one cycle to count the number of classes with consistent predictions in known class blocks. The number of different pseudo-labels activated in unlabeled blocks And reset based on the results and After initializing the model weights, retrain until the detection is successful. The trend of change reverses, that is, after a continuous increase, it decreases for the first time, or after a continuous decrease, it increases for the first time, and finally determines... and The value is .

[0103] in, Indicates the weight reduction factor. This represents the weight amplification factor; for the supervised loss weight α, if This indicates that the known classes are stable, and the weights of the supervised loss should be appropriately reduced to prevent overfitting. Let Otherwise, increase the weight of the supervision loss, making For the instance loss weight β, if This indicates that the unlabeled samples are sufficiently dispersed, and the instance loss weight should be appropriately reduced. Otherwise, increase the instance loss weight, making .

[0104] S4. Based on the loss weights determined in step S3, retrain the deep learning feature extraction network using a deep learning adaptive loss function with fixed weights until the loss function converges, and save the final trained model weight file.

[0105] Among them, the final total loss function The expression is as follows:

[0106] ;

[0107] in, This represents the supervised cross-entropy loss. Represents the instance cross-entropy loss. Indicates the contrast learning loss. and This represents the final value of the adaptively adjusted weighting coefficients.

[0108] S5. Based on step S4, use a deep learning feature extraction network and a trained model weight file to perform spatial feature inference and prediction on all data to obtain the feature vector of each sample. Then, sum the feature vectors of the samples corresponding to each IQ data and calculate the average to obtain the overall representation feature vector of the IQ data.

[0109] S6. Based on step S5, perform similarity distance calculation on the feature vectors of different IQ data to obtain the spatial relationship matrix;

[0110] Given N=100 IQ data points, we calculate the similarity distance between the feature vectors of these N IQ data points, resulting in an N*N comparison distance matrix. The ( )th element of the matrix... The value of ) is the first The IQ data and the first Spatial distance between IQ data points.

[0111] Then set the threshold as =0.9, if the spatial distance is greater than or equal to If the result is 1, then set it to 1; otherwise, set it to 0 to obtain the spatial relation matrix.

[0112] S7. Based on step S6, the spatial relationship matrix is ​​split into clusters, and the automatic aggregation of similar radiation sources is achieved by extracting connected components to obtain the radio frequency fingerprint recognition results of different IQ data.

[0113] The clustering process involves determining whether two IQ data are related to each other or can be related through other IQ files. If so, they are determined to be the same radiation source, and the same radiation sources are grouped together.

[0114] In this embodiment, the clustering effect diagram obtained by the method of the present invention is as follows: Figure 4 As shown, the clustering effect of the method of this invention in the feature space exhibits a clear category division structure. The five clusters on the left (Class 0 to Class 4) correspond to known emitter categories. The boundaries between clusters are distinct, and they are tightly clustered within each category, indicating that the supervised loss effectively anchors the decision boundaries of the known categories. The five clusters on the right (Unknown Class 1 to Unknown Class 5) correspond to automatically discovered unknown emitter categories. They not only maintain a significant distance from the known categories but are also effectively separated from each other. This proves that the method of this invention successfully disperses different unknown signals to different cluster centers through instance loss and contrastive alignment loss, and accurately aggregates them into multiple independent categories through stability-driven consensus clustering. This figure intuitively verifies the core technical effect of the method of this invention in simultaneously completing known class identification and multi-category unknown discovery in scenarios where labeled data is scarce.

[0115] In summary, the method of this invention not only has high accuracy in RF fingerprint closed-set recognition, but also has a certain recognition effect on unknown RF targets. At the same time, it has a certain effect on solving feature drift to a certain extent, realizing full automation from raw data to recognition results. Especially in scenarios where labeled data is scarce and there are unknown transmitters, it can simultaneously complete the recognition of known classes and the discovery of unknown classes. It is suitable for actual deployment environments where labeled samples are scarce and there are unknown transmitting devices, realizing high-precision recognition of known radiation sources and automatic discovery and clustering of unknown radiation sources.

[0116] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A semi-supervised radio frequency fingerprint deep clustering recognition method based on adaptive loss, the specific steps of which are as follows: S1. Obtain radio frequency signal IQ data at different time periods, perform preprocessing to obtain radio frequency training sample data, and construct a dedicated dataset. S2. Based on step S1, construct a deep learning feature extraction network and a deep learning adaptive loss function; The input to the deep learning feature extraction network is the IQ data of the radio frequency signal, and the input size is 1× The network uses blocks as the basic feature extraction modules. Each block includes a convolutional layer, a batch normalization layer, and a ReLU activation function. After feature extraction through multiple blocks, the network is finally connected to a layer containing... A fully connected layer of 10 neurons uses the softmax function to make each dimension of the vector represent a different spatial mapping position; in, This represents the preset feature space dimension; S3. Based on the deep learning feature extraction network constructed in step S2, input the special dataset constructed in step S1 to train the model and determine the model loss weights. S4. Based on the loss weights determined in step S3, retrain the deep learning feature extraction network using a deep learning adaptive loss function with fixed weights until the loss function converges, and save the final trained model weight file. Among them, the final total loss function The expression is as follows: ； in, This represents the supervised cross-entropy loss. Represents the instance cross-entropy loss. Indicates the contrast learning loss. and This represents the final value of the adaptively adjusted weighting coefficients; S5. Based on step S4, use a deep learning feature extraction network and a trained model weight file to perform spatial feature inference and prediction on all data to obtain the feature vector of each sample. Then, sum the feature vectors of the samples corresponding to each IQ data and calculate the average to obtain the overall representation feature vector of the IQ data. S6. Based on step S5, perform similarity distance calculation on the feature vectors of different IQ data to obtain the spatial relationship matrix; Given N IQ data points, calculate the similarity distance between the feature vectors of these N IQ data points to obtain an N*N comparison distance matrix; where the ( )th digit of the matrix is... The value of ) is the first The IQ data and the first Spatial distance of IQ data; Then set the threshold as If the spatial distance is greater than or equal to Set it to 1 if it is not 1, otherwise set it to 0, and obtain the spatial relation matrix; S7. Based on step S6, the spatial relationship matrix is ​​split into clusters, and the automatic aggregation of similar radiation sources is achieved by extracting connected components to obtain the radio frequency fingerprint recognition results of different IQ data. The clustering process involves determining whether two IQ data are related to each other or can be related through other IQ files. If so, they are determined to be the same radiation source, and the same radiation sources are grouped together.

2. The semi-supervised radio frequency fingerprint deep clustering recognition method according to claim 1, characterized in that, The specific steps of S1 are as follows: S11. Acquire radio frequency signal IQ data at different time periods; Each time period contains a different number of transmitters, with some transmitters appearing only in certain time periods; all signals constitute the training dataset, where a small number of samples have known class labels, denoted as the annotation set. The remaining samples are unlabeled and are referred to as the unlabeled set. Furthermore, the set is not labeled as containing unknown transmitters; let the total number of known categories be... ; S12. After preprocessing the data obtained in step S11, radio frequency training sample data is obtained. Each IQ signal is zero-mean normalized to eliminate signal amplitude differences but retain hardware fingerprint characteristics; The normalization formula is as follows: ； in, This represents the signal before normalization. The signal is normalized, and μ(x) and σ(x) represent the mean and standard deviation of the signal, respectively; and no carrier frequency offset correction, phase alignment, or IQ imbalance compensation is performed during the processing. Then based on the known number of categories And to estimate the number of possible unknown categories, and to set the dimension of the feature space. Its satisfaction Furthermore, this dimension will be used as the output dimension of the last fully connected layer of the deep learning model, and each dimension corresponds to the probability of a cluster center. in, This indicates the maximum number of unknown categories preset; Each normalized signal is then divided into continuous slices of fixed length, with a slice length of... Each slice is considered an independent sample; different slices of the same signal are considered positive sample pairs; the normalized slice samples are the radio frequency training sample data. S13. Construct a dedicated dataset based on step S12; Define the complete tag space Establish mapping This ensures that each unlabeled signal has a globally unique pseudo-label; then, the radio frequency training sample data and its corresponding labels are combined to form a dedicated dataset. in, Represents a known set of class labels. This represents a unique individual pseudo-label assigned to each unlabeled signal.

3. The semi-supervised radio frequency fingerprint deep clustering recognition method according to claim 1, characterized in that, Step S2 is as follows: First, a deep learning feature extraction network is automatically constructed based on the label space of the dedicated dataset. Then, a deep learning adaptive loss function is constructed, and the total loss function is... The weighted sum of the three losses is expressed as follows: ； in, This represents the supervised cross-entropy loss. Represents the instance cross-entropy loss. Indicates the contrast learning loss. and This represents the adaptively adjusted weight coefficients, with an initial value set to... ; For the annotation set For the samples in the dataset, the supervised cross-entropy loss is calculated using their true known class labels, as shown in the following expression: ； in, This represents the set of indices for the labeled samples in the current batch. This represents the true label of the sample. This indicates the probability that the model predicts the sample belongs to the true class. This indicates the number of labeled signals selected in the current batch. Indicates the index of the labeled signal; For unlabeled sets For the samples in step S1, the instance cross-entropy loss is calculated using the unique individual pseudo-labels assigned in step S1, as shown in the following expression: ； in, This represents the set of indices for unlabeled samples in the current batch. This represents the unique pseudo-label corresponding to the sample. This indicates the number of unlabeled signals selected in the current batch. This indicates that the signal index is not labeled; The expression for calculating the contrastive learning loss is as follows: ； in, This represents the distance loss for positive sample pairs, which forces the Euclidean distance between positive sample pairs to decrease; This represents the distance loss for negative sample pairs, which forces the Euclidean distance between negative sample pairs to increase. This represents the similarity loss of positive sample pairs, forcing the cosine similarity of positive sample pairs to increase; This represents the similarity loss of negative sample pairs, forcing the cosine similarity of negative sample pairs to decrease. The specific expressions are as follows: ； ； ； ； in, This represents a small constant to prevent overflow in logarithmic operations. The L1 norm of a matrix is ​​represented by the Euclidean distance metric. The value ranges from 0 to 1; the cosine similarity metric is... The value ranges from 0 to 1. This indicates the transpose operation. This represents a mask matrix for positive sample pairs, with dimensions N*N, where N represents the dimension of the mask matrix. With sample When from the same signal Otherwise, it is 0. Describes the mask matrix of negative sample pairs, satisfying Where 1 is an all-one matrix, and I represents the identity matrix, i.e., when And when the two samples come from different signals , and and represent the L1 norm of the positive and negative mask matrices, respectively, which is the sum of all elements in the matrix.

4. The semi-supervised radio frequency fingerprint deep clustering recognition method according to claim 1, characterized in that, Step S3 is as follows: First, set the threshold for the known class. Unknown class threshold and update factor and The deep learning feature extraction network and loss function are trained on a dedicated dataset for one cycle to count the number of classes with consistent predictions in known class blocks. The number of different pseudo-labels activated in unlabeled blocks And reset based on the results and After initializing the model weights, retrain until the detection is successful. The trend of change reverses, that is, after a continuous increase, it decreases for the first time, or after a continuous decrease, it increases for the first time, and finally determines... and The value is ; in, Indicates the weight reduction factor. This represents the weight amplification factor; for the supervised loss weight α, if This indicates that the known classes are stable, and the weights of the supervised loss should be appropriately reduced to prevent overfitting. Let Otherwise, increase the weight of the supervision loss, making For instance loss weight β, if This indicates that the unlabeled samples are sufficiently dispersed, and the instance loss weight should be appropriately reduced. Otherwise, increase the instance loss weight, making .

Images