Wireless radiation source individual identification method

Abstract

The invention relates to the technical field of wireless communication signal processing, in particular to a wireless radiation source individual identification method, which comprises the following steps of: converting an augmented sample of a received wireless signal into first signal embedding by using a signal encoder; judging a signal source through a visible radiation source identification module and generating a visible radiation source identifier; carrying out Hash coding according to signal embedding by utilizing an online signal Hash module, and generating a signal Hash code with high distinction degree; cascading the signal hash code with the seen radiation source identifier to generate a collision mitigation hash code; and predicting the individual identity tag of the radiation source on line through the index hash table, and if the collision mitigation hash code corresponding to the current signal does not exist in the hash table, considering that the signal comes from the newly added radiation source and storing the signal into the hash table for updating. Through multi-task learning, online signal processing and feature optimization technologies, the recognition accuracy and robustness of the model in an actual complex environment are remarkably improved, and the method has wide actual application value.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communication signal processing, and particularly relates to a method for identifying individual wireless radiation sources. Background Art

[0002] With the explosive growth in the number of access devices in modern wireless communication systems and the extensive expansion of application dimensions, the potential threat of malicious devices has become increasingly serious. Accurately distinguishing the individual identities of devices to ensure the security of communication systems has become a key challenge. The radiation source individual identification technology, which aims to distinguish the signals of different devices by using the inherent radio frequency fingerprints in wireless signals, has been widely applied in multiple fields such as cognitive radio, intrusion detection, and identity authentication. With the continuous progress of artificial intelligence theory and technology, the radiation source individual identification technology based on deep learning has emerged in recent years. By training a neural network model on a large-scale dataset composed of wireless received signals, it automatically extracts the identification features in the received signals and classifies them, significantly improving the ability to identify the individual identities of wireless radiation sources.

[0003] However, in practical applications, many communication devices operate in a non-cooperative manner, making it difficult to collect signal samples of all potential radiation sources on a large scale to construct a training dataset. A large number of radiation sources only provide a small amount of or no training data. These radiation sources with scarce samples can be called new radiation sources, and the identification of them can be modeled as a small-sample and zero-sample radiation source individual identification problem.

[0004] In terms of the identification problem of new radiation sources, most of the radiation source individual identification methods based on deep learning use a large number of training samples from known radiation sources to construct an identification model and identify the signals from known radiation sources in the inference stage. When encountering signals from new radiation sources, such methods can only classify all new radiation source signals into a unified unknown category. Although there is a patent describing an open-set radiation source individual identification, such a solution cannot effectively distinguish the individual identities of new radiation sources. In addition, some existing works can identify new radiation sources with only a small amount of training samples. For example, there is a patent simply describing a small-sample radiation source individual identification, but such a solution cannot effectively distinguish the individual identities of new radiation sources without training samples.

[0005] In terms of the real-time classification and identification of radiation sources, many application scenarios require quick feedback, and the model must be able to process received signals online. The existing clustering-based identification methods learn a feature extraction model through the signal samples of known radiation sources and cluster the features of the signals from new radiation sources to achieve the identification of new radiation sources. For example, there are one or two patents simply describing clustering-based zero-sample identification, but such a solution follows an offline identification paradigm and requires the signals of other new radiation sources as a reference, which cannot meet the online identification requirements.

[0006] In terms of the collision problem of signal hash coding, hash coding is an effective feature representation method that can map signals to a low-dimensional binary space. Although hash coding can initially distinguish the individual identities of radiation sources, due to the lack of sufficient training data for newly added radiation sources, existing methods are prone to overfitting, causing the model to tend to the distribution of seen radiation source data during the inference stage, resulting in a relatively high probability of hash code collisions and thus affecting classification accuracy. Summary of the Invention

[0007] Aiming at the technical bottlenecks of traditional radiation source individual recognition methods in aspects such as difficult online recognition of newly added radiation sources, serious hash code collisions, and limited generalization ability, the present invention proposes a method for wireless radiation source individual recognition, including the following steps:

[0008] Use a signal encoder to convert the augmented samples of the received wireless signals into the first signal embeddings;

[0009] Judge the signal source through the seen radiation source recognition module and generate a seen radiation source identifier;

[0010] Use the online signal hashing module to perform hash coding based on the signal embeddings to generate highly discriminative signal hash codes;

[0011] Cascade the signal hash codes and the seen radiation source identifiers to generate collision mitigation hash codes;

[0012] Online predict the radiation source individual identity label through the index hash table. If the collision mitigation hash code corresponding to the current signal does not exist in the hash table, it is considered that the signal comes from a newly added radiation source and is stored in the hash table for update.

[0013] Furthermore, during the training process, the signal encoder uses the supervised contrast loss function as the optimization objective, and optimizes the signal embeddings by maximizing the similarity between the features of the same-class samples and the differences between the features of different-class samples. The supervised contrast loss function is expressed as:

[0014]

[0015] Wherein, is the supervised contrast loss function; N is the number of original signal samples. After all the original signal samples go through two random slicing operations, each original sample will generate two augmented signal samples. Therefore, 2N is the total number of augmented signal samples; I(p) is the index set of positive samples, and |I(p)| is the cardinality of the index set I(p) of positive samples, that is, the number of samples belonging to the same class as sample p in the present invention; e p is the signal embedding of the p-th positive sample, expressed as represents the augmented signal sample corresponding to the p-th positive sample, To generate a signal embedding by using a signal encoder for and p ∈ {I(p)}; e q is the signal embedding of the q-th positive sample, where q ∈ {I(p)\p}, that is, q belongs to other samples in the index set of positive samples except the sample with index p; e k is the signal embedding of the k-th sample, where k ∈ {(I(p) ∪ I(n))\p}, that is, k is the index set of all samples except the current signal sample index p, and the index set of all samples includes the positive sample index set I(p) and the negative sample index set I(n); denotes enhancing the signal embedding e p ; Enhance the signal embedding e q ; denotes enhancing the signal embedding e k ;

[0016] Furthermore, when the identified radiation source recognition module determines the signal source, the maximum distance between the sample and the reciprocal point in the feature space is used as the decision metric to judge the sample attribution, and a radiation source identifier a is generated. When a = 1, it means the signal comes from a known radiation source, and when a = -1, it means the signal comes from a new radiation source

[0017] Furthermore, the process of determining whether the signal comes from a known radiation source includes:

[0018] Associate each known radiation source with a reciprocal point, that is, associate the radiation source of the known category u with the reciprocal point P u ;

[0019] Project the first signal embedding through a separation projector to obtain the corresponding second signal embedding;

[0020] Calculate the distance between the second signal embedding and the reciprocal point. If the distance meets the set parameters, the signal source belongs to the radiation source corresponding to the reciprocal point. If the distance between the second signal embedding and all known reciprocal points does not meet the set parameters, a new reciprocal point is added for this signal for association.

[0021] Furthermore, an adversarial reciprocal point loss is used to train the separation projector , and the adversarial reciprocal point loss is expressed as:

[0022]

[0023] where is the adversarial reciprocal point loss; is the classification error of the separation projector; is the adversarial boundary constraint of the separation projector; λ is the weight coefficient; yp is the true class label of the p-th augmented signal sample; p(u|e p ) represents the probability that the first signal embedding e of the p-th sample p belongs to class u; represents the second signal embedding z p and the point The distance between; represents the central feature point of the sample with class label y p ; R is a learnable model parameter, which in the present invention represents the boundary radius of the signal sample in the feature space, and the specific value of this parameter is obtained during the model training process.

[0024] Furthermore, the calculation process of the distance between the signal embedding and the reciprocal point includes:

[0025] d(z p ,P u ) = d s (z p ,P u ) + d a (z p ,P u )

[0026]

[0027] d a (z p ,P u ) = -z p ·P u

[0028] where d(z p ,P u ) represents the distance between the second signal embedding z p and the point P u ; d s (z p ,P u ) represents the spatial distance between the second signal embedding z p and the point P u ; n represents the dimension of the embedding space. In the present invention, is used as a normalization factor to prevent the magnitude of the distance value from being too large in different dimensions, thereby improving the calculation stability; ||·|| 2 represents the calculation of the L2 norm; d a (z p ,P u ) represents the angular distance between the second signal embedding z p and the point P u .

[0029] Furthermore, a supervised contrastive loss function is used to train the online signal hashing module, so that when using the online signal hashing module to perform hashing encoding based on signal embeddings, the hashing encodings of signals in the same category are close, and the hashing codes of signals in different categories are different. The supervised contrastive loss function is expressed as:

[0030]

[0031] Wherein, is the supervised contrastive loss function; z' p is the embedding feature vector of the p-th sample, expressed as h p represents the hashing code generated by the online signal hashing module according to the p-th sample, and c p represents the confidence level generated by the online signal hashing module according to the p-th sample. represents the Kronecker product; z' q represents the embedding feature vector of the q-th positive sample; z' k is the embedding feature vector of the k-th sample; z' p ·z' k represents the inner product operation between the vector z' p and the vector z' k to calculate the similarity between embedding feature vectors.

[0032] Furthermore, an embedding enhancer is used to perform data enhancement on the first signal embedding, and the enhanced data is used as the input of the online signal hashing module.

[0033] Furthermore, a binary constraint is introduced for the signal hashing code. The binary constraint is used to ensure that the generated signal hashing code is close to the binary code, so as to improve the efficiency and stability of model training. The binary constraint regularization term is expressed as:

[0034]

[0035] Wherein, is the binary constraint regularization term; ||h p || represents the calculation of the L1 norm of the hashing code h p ; L represents the length of the hashing code, that is, how many binary bits are included in the hashing code of each sample. In the present invention, this parameter is a learnable model parameter, and the optimal value of this parameter is obtained through the training process. Those skilled in the art can also set it according to experience.

[0036] Furthermore, a similarity constraint is introduced for the signal hashing code. The similarity constraint ensures that the hashing codes of signals from the same radiation source have high similarity, while the hashing codes of signals from different radiation sources are separated as much as possible. The similarity constraint regularization term is expressed as:

[0037]

[0038] Among them,

[0039] is the similarity constraint regular term; h q represents the hash code generated by the online signal hashing module according to the q-th sample; h k represents the hash code generated by the online signal hashing module according to the k-th sample; h p ·h q represents p the inner product operation of h q and h

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. The present invention adopts an online radiation source individual recognition method based on signal hashing. By learning general knowledge from seen transmitters, discriminative hash codes are assigned to signals from different transmitters and a hash table is indexed for classification decision-making, providing a unified method for online recognition of new radiation source individuals that is applicable to both small-sample and zero-sample scenarios, and capable of online feedback to meet the requirements of efficient and accurate radiation source individual identity recognition in practical applications;

[0042] 2. The present invention designs a collision mitigation signal hashing model. The model uses a signal encoding module to extract discriminative embedding features and a seen radiation source recognition module to generate a seen radiation source identifier indicating whether the signal is from a seen radiation source, which is cascaded to the signal hash code output by the online signal hashing module to generate a collision mitigation hash code, thereby reducing the hash code collision probability between seen radiation sources and new radiation sources and improving the recognition accuracy of new radiation sources, enabling the model to still efficiently recognize signals from new radiation sources in the absence of training data for new radiation sources;

[0043] 3. The present invention adopts a multi-task learning strategy during the training process, fuses multiple tasks such as embedding feature extraction, hash coding optimization, and seen radiation source recognition to design a loss function, and further enhances the generalization ability of the embedding features through the joint optimization of the adversarial reciprocal point loss and the supervised contrast loss, enabling the model to adapt to different channel environments and various signal systems and improving the overall recognition performance.

[0044] In summary, in view of the technical bottlenecks of traditional radiation source individual recognition methods in aspects such as difficult online recognition of new radiation sources, serious hash code collisions, and limited generalization ability, the present invention proposes an improved solution based on the Collision-Aware Signal Hashing (CASH) model. Through multi-task learning, online signal processing, and feature optimization technologies, the present invention significantly improves the recognition accuracy and robustness of the model in actual complex environments, and has broad practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic flow chart of a method for individual recognition of wireless radiation sources according to the present invention;

[0046] Figure 2 It is a module architecture diagram of a method for individual recognition of wireless radiation sources according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0048] The present invention proposes a method for individual recognition of wireless radiation sources, as Figure 1 , including the following steps:

[0049] Use a signal encoder to convert the augmented sample of the received wireless signal into a first signal embedding;

[0050] Judge the signal source through the known radiation source recognition module and generate a known radiation source identifier;

[0051] Use the online signal hashing module to perform hash encoding based on the signal embedding to generate a highly discriminative signal hash code;

[0052] Cascade the signal hash code and the known radiation source identifier to generate a collision-aware hash code;

[0053] Online predict the individual identity label of the radiation source through the index hash table. If the collision-aware hash code corresponding to the current signal does not exist in the hash table, it is considered that the signal comes from a new radiation source and is stored in the hash table for update.

[0054] In this embodiment, random slicing in common augmentation methods is used to obtain augmented samples. During model training, each signal sample is randomly sliced twice to generate two augmented samples from the original signal sample. The signal augmented samples from the same radiation source are positive samples, and the signal augmented samples from different radiation sources are negative samples, thereby improving the translational invariance of signal embedding. During recognition, an augmented sample of the same length is sliced out from the center of each signal sample. The signal encoder extracts signal embeddings with strong generalization ability and class discrimination ability to support subsequent hash coding and recognition tasks. The signal encoder consists of an embedding extractor and an embedding enhancer. The embedding extractor generates a signal representation related to the radio frequency fingerprint as the signal embedding; the embedding enhancer further maps the signal embedding into low-dimensional features to enhance the generalization ability of the signal embedding.

[0055] During the model training process, the signal encoder mainly uses the supervised contrast loss function as the optimization objective. By maximizing the similarity between the features of samples of the same class and the difference between the features of samples of different classes, the signal embedding is optimized. The supervised contrast loss function is:

[0056]

[0057] where is the supervised contrast loss function; N is the number of original signal samples. After all the original signal samples go through two random slicing operations, each original sample will generate two augmented signal samples. Therefore, 2N is the total number of augmented signal samples; I(p) is the index set of positive samples, and |I(p)| is the cardinality of the index set I(p) of positive samples, that is, the number of samples belonging to the same class as sample p; e p is the signal embedding of the p-th positive sample, denoted as represents the augmented signal sample corresponding to the p-th positive sample, is to use the signal encoder to map to generate a signal embedding; e q is the signal embedding of the q-th positive sample, q is the index of the signal embedding of the positive sample belonging to the same class as e p , and q comes from the positive sample index set I(p); e k is the signal embedding of the k-th sample, and k is the index set of all samples except the current signal sample index p, including positive sample indices and negative sample indices; represents enhancing the signal embedding e p ; enhances the signal embedding e q ; represents enhancing the signal embedding e k .

[0058] In this embodiment, the seen radiation source identification module generates a seen radiation source identifier, which reflects whether the received signal belongs to any seen radiation source. The present invention uses the adversarial reciprocal point loss to guide the model to learn a feature space where the seen classes and the new classes can be clearly separated, making the embeddings of the same-class signal samples more compact and the embeddings of the new-class signals clearly separated from them. When identifying, the maximum distance between the sample and the reciprocal point in the feature space is used as the decision metric to judge the sample's belonging, and a radiation source identifier a is generated, where a = 1 indicates that the signal comes from a seen radiation source, and a = -1 indicates that the signal comes from a new radiation source. The identification process includes:

[0059] Associate a reciprocal point with each known radiation source, that is, associate the radiation source of the known class u with the reciprocal point P u ;

[0060] Project the first signal embedding through the separation projector to obtain the corresponding second signal embedding;

[0061] Calculate the distance between the second signal embedding and the reciprocal point. If the distance meets the set parameters, the signal source belongs to the radiation source corresponding to the reciprocal point. If the distances between the second signal embedding and all the known reciprocal points do not meet the set parameters, a new reciprocal point is added for this signal for association.

[0062] The distance between the signal embedding and the class reciprocal point can be divided into a spatial distance and an angular distance, specifically as follows:

[0063] d(z p ,P u ) = d s (z p ,P u ) + d a (z p ,P u )

[0064]

[0065] d a (z p ,P u ) = -z p ·P u

[0066] where d(z p ,P u ) represents the distance between the second signal embedding z p and the point P u ; d s (z p ,P u ) represents the distance between the second signal embedding z p and the point P uThe spatial distance between them; n represents the dimension of the embedding space; Represents the normalization factor to prevent the magnitude of distance values ​​in different dimensions from being too large, thereby improving the calculation stability; ||·|| 2 Indicates the calculation of L2 norm; d a (z p ,P u ) represents the second signal embedded in z p With point P u The angular distance between .

[0067] To ensure the compactness of samples of seen categories, the adversarial reciprocity point loss includes the classification error part:

[0068]

[0069] Among them, y p is the category label of the pth augmented signal sample; p(u|e p ) represents the first signal embedding e of the pth sample p The probability of belonging to category u is specifically expressed as:

[0070]

[0071] in, It represents the set of all categories, including the categories of radiation sources that have been seen and the categories of new radiation sources; d(z p ,P v ) represents the second signal embedded in z p With point P v The distance between v Represents the reciprocity point corresponding to category v.

[0072] In addition, the present invention introduces adversarial boundary constraints to ensure reliable separation of new radiation source signals, which is specifically expressed as:

[0073]

[0074] in, The second signal is embedded in z p With point The distance between Indicates that the category label is y p The characteristic center point of the sample; R is a learnable model parameter, which represents the boundary radius of the signal sample in the feature space.

[0075] Finally, the adversarial reciprocity point loss can be expressed as:

[0076]

[0077] Among them, λ is a weight coefficient used to balance the weight coefficient of the adversarial constraint strength.

[0078] To reduce the training complexity, in this embodiment, the signal encoder module and the seen emitter recognition module are trained simultaneously using a multi-task training mode, thereby enhancing the generalization ability of signal embedding, reducing the training cost at the same time, and finally the loss in the first stage can be expressed as:

[0079]

[0080] Among them, is the supervised contrast loss function of the signal encoding module; α is an adjustable weight used to balance and the relative importance between; is the final adversarial reciprocal point loss of the seen emitter recognition module.

[0081] In this embodiment, if multiple signal samples come from the same emitter, they will be assigned the same hash code; signal samples from different emitters will be assigned different hash codes. Specifically, the training process of the online signal hashing module includes:

[0082] The signal embedding is processed by a sign projector and a confidence projector;

[0083] The signal hash code is optimized through the supervised contrast loss to ensure that the generated hash code has strong discriminability;

[0084] Model regularization is performed by combining binary constraints and similarity constraints to improve the discriminability of the hash code.

[0085] Specifically, the signal embedding is processed by a sign projector and a confidence projector. The sign projector generates an attribute identifier as the signal hash code, and each dimension represents whether the signal has a certain emitter attribute. The confidence projector calculates the confidence in the signal attribute decision based on the signal embedding. To ensure that the generated hash code has strong discriminability, the present invention uses the supervised contrast loss to optimize the signal hash code. The purpose of this loss function is to maximize the similarity of the signal hash codes from the same emitter and minimize the similarity of the signal hash codes from different emitters. The supervised contrast loss optimizes the relative distance of the signal hash code, making the hash codes of similar signals as close as possible and the hash codes of different signals as far apart as possible. The specific supervised contrast loss function can be expressed as:

[0086]

[0087] Among them, is the supervised contrast loss function; N is the number of original signal samples. After two random slicing operations, each original sample will generate two augmented signal samples, so 2N is the total number of augmented signal samples; I(p) is the index set of positive samples, |I(p)| is the cardinality of the index set I(p) of positive samples, that is, the number of samples belonging to the same category as sample p; z' p is the embedded feature vector of the pth sample, expressed as h p represents the hash code generated by the online signal hash module according to the pth sample, c p represents the confidence generated by the online signal hash module based on the p-th sample, represents the Kronecker product, which combines the hash code and confidence to improve the discrimination ability; q is the p The positive sample signals belonging to the same category are embedded into indexes, and q comes from the positive sample index set I(p); z' q represents the embedded feature vector of the qth positive sample; k is the index set of all samples except the current signal sample index p, including positive sample index and negative sample index; z' k is the embedded feature vector of the kth sample; represents the inner product operation, which is used to calculate the similarity between embedded feature vectors.

[0088] In order to further improve the distinguishability of hash codes, the present invention additionally introduces binary constraints and similarity constraints for signal hash codes. Binary constraints are used to ensure that the generated signal hash codes are close to binary codes to improve the efficiency and stability of model training. Similarity constraints ensure that the signal hash codes from the same radiation source have high similarity, and the signal hash codes from different radiation sources are separated as much as possible. The specific binary constraint and similarity constraint regularization terms are expressed as follows:

[0089]

[0090]

[0091] in, is a binary constraint regularization term; ||h p || indicates the L1 norm calculation of the hash code; L is an adjustable hyperparameter, which indicates the length of the hash code, that is, how many binary bits are contained in the hash code of each sample. is the similarity constraint regularization term; h q represents the hash code generated by the online signal hash module according to the qth sample; h k represents the hash code generated by the online signal hash module according to the kth sample; represents the inner product operation, which is used to calculate the similarity between embedded feature vectors.

[0092] To ensure that the signal hash codes from different emission sources have good discrimination ability, in this embodiment, the binary constraint regularization term and the similarity constraint regularization term are merged and incorporated into the regularization term of the online signal hashing module to effectively avoid the problem of hash code collision and improve the robustness and accuracy of the model when facing new signal sources. Specifically, it can be expressed as:

[0093]

[0094] Combined with the regularization formed by the symbol projector, confidence projector and double constraints of the online signal hashing module, the present invention significantly reduces the influence of intra-class perturbation, improves the discrimination of the signal hash code, and thus enhances the accuracy and robustness of the model. The loss function of the final second stage can be expressed as:

[0095]

[0096] Among them, β is an adjustable weight used to balance the supervised contrast loss and the regularization term, so as to determine whether to focus more on learning the discriminative features of the signal or strengthening the generalization ability of the model during the optimization process.

[0097] Cascade the signal hash code with the identifier of the seen radiation source to generate a collision mitigation hash code, which can be specifically expressed as:

[0098] h c =[h, a];

[0099] The collision mitigation hash code not only contains the feature information of the signal, but also contains the additional information on whether the signal comes from a seen radiation source.

[0100] When multiple signal samples come from the same radiation source, their hash codes should be similar, and the collision mitigation hash codes after splicing can still effectively distinguish them. However, if signals from different radiation sources have similar hash codes, the identifier will effectively avoid the hash code collision problem, ensuring that even if the hash codes are the same, the source of the signal can still be distinguished based on the identifier.

[0101] In this embodiment, the collision mitigation hash code is indexed through a hash table. When a new signal sample arrives, we will check whether the current hash code already exists in the hash table S: if it does not exist in the hash table S, it means that the signal comes from a new radiation source, and we will add it as a new item to the hash table and perform label prediction on the signal; if it already exists in the hash table, it means that the signal comes from a seen radiation source, and the hash table will return the stored identity label. Specifically, it is expressed as follows:

[0102]

[0103] Among them, Ind(S, h c ) is the tag corresponding to the collision mitigation hash code h c in the hash table S. If h c is not in the hash table, the model will regard it as a new radiation source and update it.

[0104] According to the above complete process of the present invention,

[0105] This embodiment will illustrate the training process and the real-time data processing process in the present invention. As Figure 2 , the training process specifically includes:

[0106] Signal samples obtained from different radiation sources are processed based on the signal augmentation technique to obtain augmented samples;

[0107] The embedding extractor of the signal encoder obtains the first signal embedding from the augmented samples, and the embedding enhancer of the signal encoder performs data augmentation on the first signal embedding. Using the supervised contrast loss function to train the signal encoder;

[0108] The separation projector of the known radiation source recognition module obtains the corresponding second signal embedding according to the input first signal embedding, inputs the second signal embedding into a classifier or other decision modules to determine whether the category of the signal belongs to a known category, and marks it with an identifier. The known radiation source recognition module is trained according to the classification error and the adversarial boundary constraint;

[0109] Taking the output of the embedding enhancer as the input of the online signal hashing module, using the confidence projector and the sign projector to obtain the confidence and hash code of the input signal respectively, and training the online signal hashing module with the supervised contrast loss function combined with the binary constraint and the similarity constraint.

[0110] As Figure 2 , the real-time data processing process specifically includes:

[0111] Signal samples obtained from different radiation sources are processed based on the signal augmentation technique to obtain augmented samples;

[0112] The pre-trained embedding extractor obtains the first signal embedding from the augmented samples, and the pre-trained embedding enhancer performs data augmentation on the first signal embedding;

[0113] The known radiation source recognition module marks the signal with an identifier according to the first signal embedding. When the signal comes from a known radiation source, the identifier a = 1; when the signal comes from an unknown radiation source, the identifier a = -1;

[0114] Taking the output of the embedding enhancer as the input of the sign projector in the online signal hashing module to obtain the hash code corresponding to the signal;

[0115] Concatenate the identifier and hash code of the signal as the collision mitigation hash code of the signal.

[0116] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying an individual wireless radiation source, characterized in that: The following steps are involved: converting the augmented samples of the received wireless signal into a first signal embedding using a signal encoder; Determine the signal source and generate a seen radiation source identifier by using a seen radiation source identification module; Use the online signal hash module to perform hash coding based on signal embedding to generate a highly discriminative signal hash code; concatenating the signal hash code with the seen radiator identifier to generate a collision mitigation hash code; The individual identity labels of the radiation sources are predicted online by indexing the hash table. If the collision mitigation hash code corresponding to the current signal does not exist in the hash table, the signal is considered to come from a newly added radiation source and is stored in the hash table for updating.

2. A method for identifying an individual wireless radiation source according to claim 1, characterized in that: During the training process, the signal encoder uses the supervised contrast loss function as the optimization target to optimize signal embedding by maximizing the similarity between features of samples of the same type and the difference between features of samples of different types. The supervised contrast loss function is expressed as: in, is the supervised contrast loss function; N is the number of original signal samples; I(p) is the index set of positive samples, |I(p)| is the cardinality of the index set I(p) of positive samples; e p is the signal embedding of the pth positive sample, expressed as represents the augmented signal sample corresponding to the pth positive sample, To utilize the signal encoder Mapping is performed to generate signal embedding, p∈{I(p)}; e q is the signal embedding of the qth positive sample, q∈{I(p)\p}; e k is the signal embedding of the kth sample, k∈{(I(p)∪I(n))\p}, I(n) is the index set of negative samples; Indicates the signal embedding e p Perform enhancement processing; Embed the signal q Perform enhancement processing; Indicates the signal embedding e k Perform enhancement processing.

3. A method for identifying an individual wireless radiation source according to claim 1, characterized in that: It has been seen that when the radiation source identification module determines the source of the signal, the maximum distance of the sample from the reciprocity point in the feature space is used as the decision metric to determine the sample attribution and generate a radiation source identifier a. When a=1, it means that the signal comes from a known radiation source, and when a=-1, it means that the signal comes from a newly added radiation source.

4. A method for identifying an individual wireless radiation source according to claim 3, characterized in that: The process of determining whether a signal comes from a known radiation source includes: Associating each known radiation source with a reciprocity point, that is, associating the radiation source of known category u with the reciprocity point P u ; The first signal is embedded through a separate projector Performing projection to obtain a corresponding second signal embedding; The distance between the second signal embedding and the reciprocity point is calculated. If the distance meets the set parameters, the signal source belongs to the radiation source corresponding to the reciprocity point. If the distance between the second signal embedding and all known reciprocity points does not meet the set parameters, a new reciprocity point is added to the signal for association.

5. A method for identifying an individual wireless radiation source according to claim 4, characterized in that: Adopting countermeasure reciprocity point loss to separate projectors For training, the adversarial point loss is expressed as: in, To combat reciprocity point loss; To separate the classification error of the projector; is the adversarial boundary constraint of the separation projector; λ is the weight coefficient; N is the number of original signal samples; y p is the true category label of the pth augmented signal sample; p(u|e p ) represents the first signal embedding e of the pth sample p The probability of belonging to category u; The second signal is embedded in z p With point The distance between Indicates that the category label is y p The characteristic center point of the sample; R is the learnable model parameter.

6. A method for identifying an individual wireless radiation source according to claim 4 or 5, characterized in that: The calculation process of the distance between the signal embedding and the reciprocity point includes: d(z p ,P u )=d s (z p ,P u )+d a (z p ,P u ) d a (z p ,P u )=-z p ·P u Among them, d(z p ,P u ) represents the second signal embedded in z p With point P u The distance between s (z p ,P u ) represents the second signal embedded in z p With point P u The spatial distance between them; n represents the dimension of the embedding space; ||·||2 represents the calculation of the L2 norm; d a (z p ,P u ) represents the second signal embedded in z p With point P u The angular distance between .

7. A method for identifying individual wireless radiation sources according to claim 1, characterized in that: The supervised contrast loss function is used to train the online signal hash module, so that when the online signal hash module is used to perform hash coding according to signal embedding, the hash codes of signals of the same category are close and the hash codes of signals of different categories are different. The supervised contrast loss function is expressed as: in, is the supervised contrast loss function; N is the number of original signal samples; I(p) is the index set of positive samples, |I(p)| is the cardinality of the index set I(p) of positive samples; z' p is the embedded feature vector of the pth sample, expressed as h p represents the hash code generated by the online signal hash module according to the pth sample, c p represents the confidence generated by the online signal hash module based on the p-th sample, represents the Kronecker product, p∈{I(p)}; z' q Represents the embedded feature vector of the qth positive sample, q∈{I(p)\p}; z' k is the embedded feature vector of the kth sample, k∈{(I(p)∪I(n))\p}, I(n) is the index set of negative samples; z' p ·z' k Represents vector z' p With vector z' k The inner product operation between .

8. A method for identifying individual wireless radiation sources according to claim 7, characterized in that: The embedding enhancer is used to perform data enhancement on the first signal embedding, and the enhanced data is used as input to the online signal hashing module.

9. A method for identifying individual wireless radiation sources according to claim 7, characterized in that: Binary constraints are introduced for signal hash codes. Binary constraints are used to ensure that the generated signal hash codes are close to binary codes to improve the efficiency and stability of model training. The binary constraint regularization term is expressed as: Among them, c BIN is a binary constraint regularization term; h p represents the hash code generated by the online signal hash module according to the pth sample, ||h p || represents the hash code h p Perform L1 norm calculation; L represents the length of the hash code.

10. A method for identifying an individual wireless radiation source according to claim 7 or 9, characterized in that: A similarity constraint is introduced for the signal hash code. The similarity constraint ensures that the signal hash codes from the same radiation source have high similarity, and the signal hash codes from different radiation sources are separated as much as possible. The similarity constraint regularization term is expressed as: Among them, c SIM is the similarity constraint regularization term; h q represents the hash code generated by the online signal hash module according to the qth sample; h k represents the hash code generated by the online signal hash module according to the kth sample; h p ·h q Indicates h p With h q Inner product operation.

Images