CN-121997002-A - Transformer encoder network-based cross-channel radiation source individual identification method

CN121997002ACN 121997002 ACN121997002 ACN 121997002ACN-121997002-A

Abstract

The invention provides a trans-former encoder network-based cross-channel radiation source individual identification method, and belongs to the technical field of information sensing and identification. The method comprises the steps of collecting radiation source sampling signals under the conditions of a source domain transmission channel and a target domain transmission channel, generating a differential constellation track diagram through data processing, constructing a cross-channel radiation source individual identification sample data set, constructing a domain self-adaptive countermeasure network model, establishing a cross entropy loss function, simultaneously completing countermeasure training of a channel type discrimination branch and an individual label prediction branch in the domain self-adaptive countermeasure network model by adopting an Adam optimization algorithm, processing a target domain transmission channel signal waveform to form the differential constellation track diagram, predicting a radiation source individual identification label by utilizing the trained domain self-adaptive countermeasure network model, and completing radiation source individual identification of a target domain transmission channel received signal waveform. The method can rapidly and accurately identify the radiation source signal individuals under different transmission channel types.

Inventors

ZHANG GUANJIE
LU NINGNING
HU YANG
RONG QIANG
LI YANBIN
Jiang menglan
CHEN TAOYI
CHEN JINYONG
QIAO QIANG
SHI TUO
LI CHUNZE
LIU CHUNRAN

Assignees

中国电子科技集团公司第五十四研究所
中电网络空间研究院有限公司

Dates

Publication Date: 20260508
Application Date: 20260403

Claims (5)

1. The trans-former encoder network-based cross-channel radiation source individual identification method is characterized by comprising the following steps of: Step 1, collecting radiation source sampling signals under the conditions of a source domain transmission channel and a target domain transmission channel, generating a differential constellation track diagram through data processing, and constructing a cross-channel radiation source individual identification sample data set; step2, constructing a domain self-adaptive countermeasure network model; step 3, establishing a cross entropy loss function, and simultaneously completing countermeasure training for a channel type discrimination branch and an individual label prediction branch in a domain self-adaptive countermeasure network model by adopting an Adam optimization algorithm; And 4, processing the signal waveform of the target domain transmission channel to form a differential constellation track diagram, and predicting the radiation source individual identification tag by using the trained domain self-adaptive countermeasure network model to finish the radiation source individual identification of the signal waveform received by the target domain transmission channel.
2. The method for identifying the individual cross-channel radiation source based on the transducer encoder network according to claim 1, wherein the specific mode of the step 1 is as follows: (101) Performing frequency estimation under a non-cooperative mode on an original sampling signal under a source domain transmission channel, and then performing down-conversion, truncation, normalization, over-sampling and differential processing to form a differential constellation track diagram under the source domain transmission channel, and annotating radiation source individual information and transmission channel types on the differential constellation track diagram of the source domain transmission channel to form source domain signal sample data; (102) Performing frequency estimation under a non-cooperative mode on an original sampling signal under a target domain transmission channel, and then performing down-conversion, truncation, normalization, over-sampling and differential processing to form a differential constellation track diagram under the target domain transmission channel, and marking the differential constellation track diagram of the target domain transmission channel with only the type of the transmission channel to form target domain signal sample data; (103) The source domain signal samples and the target domain signal samples together form a cross-channel radiation source individual identification sample dataset.
3. The method for identifying the cross-channel radiation source individuals based on the transponder encoder network according to claim 1, wherein the domain self-adaptive countermeasure network model comprises a transponder encoder network module, a gradient turnover layer, a channel type distinguishing network module and an individual tag predicting network module, wherein the transponder encoder network module is used for extracting fingerprint characteristics of signal samples, the gradient turnover layer and the channel type distinguishing network module are connected in series to form a channel type distinguishing branch, the individual tag predicting network module is used as a radiation source individual tag distinguishing branch, and the channel type distinguishing branch and the radiation source individual tag predicting branch are connected in parallel behind the transponder encoder network module; The transducer encoder network consists of 1 layer of convolution layer and 3 layers of encoder network modules, the convolution layer maps the differential constellation locus diagram into 1 embedded vector, the embedded vector and locus diagram position vector form the encoder network module together, and each layer of encoder network module consists of 8 multi-head attention layers, a forward connection layer and layer normalization; the channel type judging network and the individual label predicting network are all composed of a full-connection layer of 2 layers; the gradient inversion layer is a neural network that inverts the direction of the gradient during back propagation.
4. The method for identifying the cross-channel radiation source individual based on the transducer encoder network according to claim 1, wherein in the step 3, the maximization of the channel type discrimination error is realized while the individual tag prediction error is minimized, so that the feature vectors of the source domain signal sample and the target domain signal sample extracted by the transducer encoder network are distributed consistently in the feature space, and the specific manner of the step 3 is as follows: (301) Establishing an individual label prediction multi-classification cross entropy loss function: Wherein, the Representing the predicted value of the tag predictor, Representing a real radiation source tag, Representing the number of radiation sources, subscripts Is the sample sequence number; (302) For known parameters of source domain transmission channel signal samples and target domain transmission channel signal samples, establishing a channel type discrimination two-class cross entropy loss function: Wherein, the Representing the predicted value of the domain discrimination module, Representing real domain labels, subscripts Representing a sample number; (303) Establishing a total loss function of a domain adaptive neural network model: Wherein, the Representing the number of source domain signal samples, Representing the number of samples of the signal in the target domain, Represents the domain discrimination loss of the source domain signal samples, A domain discrimination loss representing a target domain signal sample; (304) And carrying out repeated iterative training on the domain self-adaptive countermeasure network model by adopting an Adam optimization algorithm until network parameters tend to be stable.
5. The method for identifying the individual cross-channel radiation source based on the transducer encoder network according to claim 1, wherein the specific mode of the step 4 is as follows: (401) Processing the target domain transmission channel signal sample to be predicted to generate a differential constellation track diagram; (402) And predicting the individual numbers of the radiation sources of the signal samples collected by the target domain transmission channel by using the domain self-adaptive countermeasure network model after training, and completing the individual identification of the radiation sources of the signal samples of the target domain transmission channel.

Description

Transformer encoder network-based cross-channel radiation source individual identification method Technical Field The invention belongs to the technical field of information sensing and identification, and particularly relates to a trans-former encoder network-based cross-channel radiation source individual identification method. Background The specific radiation source identification is an identification technology for extracting radio frequency fingerprint characteristics from received signals to distinguish individuals with different radiation sources of the same type, and can effectively improve the capabilities of electromagnetic spectrum sensing, illegal user intrusion detection and the like. The current radiation source individual identification method is mainly focused on the condition that source domain signals and target domain signal samples to be identified are distributed in the same mode, radio frequency fingerprint features are extracted by means of manual design, distortion signal models or neural network models and the like, and the identification is carried out by combining classifiers of different architectures. Along with the diversification of individual identification scenes of radiation sources, multipath effects, doppler effects and other factors of different transmission channels can change the fingerprint feature distribution of received signals extracted by the existing method, the performances of a fingerprint feature extraction model and a classifier model constructed under the condition of a single channel (the source domain and the target domain are identical in distribution) are drastically reduced in the scenes of different transmission channels, second, received signal samples after channel change often have no individual tag information, the fingerprint feature extraction model and the classifier model cannot be retrained by directly utilizing the signal samples after channel change, and finally, the existing feature extraction method has the phenomenon of strong dependence on signal processing expertise. Aiming at the defects of radiation source individual identification caused by channel variation, the invention provides a domain adaptive neural network identification method based on DRSN-transducer encoder network. Disclosure of Invention Aiming at the defects of the existing cross-channel radiation source individual identification method, the invention provides a cross-channel radiation source individual identification method based on a transducer encoder network. The method can rapidly and accurately identify the radiation source signal individuals under different transmission channel types. The invention adopts the technical scheme that: A trans-former encoder network-based cross-channel radiation source individual identification method comprises the following steps: Step 1, collecting radiation source sampling signals under the conditions of a source domain transmission channel and a target domain transmission channel, generating a differential constellation track diagram through data processing, and constructing a cross-channel radiation source individual identification sample data set; step2, constructing a domain self-adaptive countermeasure network model; step 3, establishing a cross entropy loss function, and simultaneously completing countermeasure training for a channel type discrimination branch and an individual label prediction branch in a domain self-adaptive countermeasure network model by adopting an Adam optimization algorithm; And 4, processing the signal waveform of the target domain transmission channel to form a differential constellation track diagram, and predicting the radiation source individual identification tag by using the trained domain self-adaptive countermeasure network model to finish the radiation source individual identification of the signal waveform received by the target domain transmission channel. Further, the specific mode of the step 1 is as follows: (101) Performing frequency estimation under a non-cooperative mode on an original sampling signal under a source domain transmission channel, and then performing down-conversion, truncation, normalization, over-sampling and differential processing to form a differential constellation track diagram under the source domain transmission channel, and annotating radiation source individual information and transmission channel types on the differential constellation track diagram of the source domain transmission channel to form source domain signal sample data; (102) Performing frequency estimation under a non-cooperative mode on an original sampling signal under a target domain transmission channel, and then performing down-conversion, truncation, normalization, over-sampling and differential processing to form a differential constellation track diagram under the target domain transmission channel, and marking the differential constellation track diagram of the target domain transmiss