Search

CN-116738209-B - Radiation source radio frequency characteristic extraction method based on measurement and deep learning

CN116738209BCN 116738209 BCN116738209 BCN 116738209BCN-116738209-B

Abstract

A radiation source radio frequency characteristic extraction method based on measurement and deep learning belongs to the technical field of radiation source radio frequency characteristic extraction. The invention solves the problems that the RF characteristics with good stability and separability are difficult to extract by adopting the prior method, and the extracted RF characteristics are easy to be interfered by IM information to fail. The method comprises the steps of constructing a modeling data set and an ideal training data set, processing each RF signal sample in the modeling data set to obtain a processing result corresponding to each RF signal sample, training a built AE network by using the processing result of the RF signal sample in the ideal training data set, training a FRM network by using the processing result of the RF signal sample in the modeling data set and the trained AE network, processing RF signals to be detected, and inputting the processing result into the trained FRM network constrained by the output of a signal feature encoder to obtain a radiation source radio frequency feature extraction result. The invention can be applied to the radio frequency characteristic extraction of the radiation source.

Inventors

  • LIU LUTAO
  • ZHANG WEI
  • JIANG YILIN

Assignees

  • 哈尔滨工程大学

Dates

Publication Date
20260508
Application Date
20230621

Claims (10)

  1. 1. The method for extracting the radio frequency characteristics of the radiation source based on measurement and deep learning is characterized by comprising the following steps of: Step 1, modeling N types of radiation source structure models, respectively generating P radiation source individuals for each type of radiation source structure model by adjusting UM parameters, wherein UM is derived from electronic components in the radiation source and is added to an IM signal in a UM mode; generating an RF signal sample set carrying RF information by adjusting IM modulation parameters of each radiation source individual, and taking the generated RF signal sample set carrying RF information as a modeling data set; Generating an RF signal sample set which does not carry RF information according to IM modulation parameters adopted when generating a modeling data set, and taking the generated RF signal sample set which does not carry RF information as an ideal training data set; Step 2, processing each RF signal sample in the modeling data set and the ideal training data set respectively to obtain a processing result corresponding to each RF signal sample; Step 3, building an AE network, wherein the AE network comprises a signal characteristic encoder And a reconstruction decoder Training the constructed AE network by using the processing result of the RF signal sample in the ideal training data set; step 4, designing a FRM network, wherein the FRM network comprises an RF feature encoder Reconstruction decoder Metric score device Training the FRM network by using the processing result of the RF signal sample in the modeling data set and the trained AE network; the input of the RF feature encoder is an RF signal carrying RF information; the input of the reconstruction decoder is the splicing result of the depth RF characteristics and the depth signal characteristics, specifically, the RF signals carrying the RF information are respectively input into an RF characteristic encoder in the FRM network and a signal characteristic code in the trained AE network so as to automatically learn the depth RF characteristics and extract the depth signal characteristics; the metric score calculates a similarity score between each RF feature pair; Step 5, acquiring an RF signal to be detected, processing the acquired signal, and inputting the processing result into a trained signal feature encoder FRM network with output constraint, RF feature encoder As a result of the radiation source rf signature extraction.
  2. 2. The method for extracting RF characteristics of a radiation source based on metric and deep learning of claim 1, wherein the RF signal samples not carrying RF information are: Wherein, the For RF signal samples that do not carry RF information, As the baseband signal, the signal is a signal, Is an IM system; the RF signal sample carrying RF information is: Wherein, the In order to carry RF signal samples of RF information, Is an RF system.
  3. 3. The method for extracting RF characteristics of a radiation source based on metric and deep learning as claimed in claim 2, wherein the processing manner of the RF signal sample is: For any RF signal sample, extracting in-phase data and quadrature data from the RF signal sample by concatenating in-phase data and quadrature data: Wherein, the Is the data of the same phase, As the data to be orthogonal to each other, Is a serial result; For series results And carrying out maximum absolute value normalization processing, wherein the processing result corresponding to the RF signal sample is as follows: Wherein, the For the processing result corresponding to the RF signal sample, Is the maximum absolute value of the data in the series result.
  4. 4. The method for extracting radio frequency characteristics of a radiation source based on measurement and deep learning according to claim 3, wherein an optimizer adopted in training the built AE network is an adaptive moment estimation optimizer, the learning rate is set to be 0.0001, and an adopted objective function is as follows: Wherein, the To train the objective function employed in the AE network, Is the first of ideal training data set The processing results corresponding to the individual RF signal samples, Is that Through the output of the AE network, The number of the phases is 2-norm, Representing the batch size.
  5. 5. The method for extracting RF characteristics from a radiation source based on metric and deep learning as claimed in claim 4, wherein said method comprises the steps of The output through the AE network is: Wherein, the In the case of a signal feature encoder, As a parameter of the signal characteristic encoder, For the output of the signal characteristic encoder, A reconstruction of the decoder is performed such that, Parameters of the reconstructed decoder.
  6. 6. The method for extracting radio frequency characteristics of a radiation source based on measurement and deep learning as set forth in claim 5, wherein the optimizer adopted in the training of the FRM network is an adaptive moment estimation optimizer, the learning rate is set to 0.0001, and the adopted objective function is: Wherein, the To train the objective function employed in the FRM network, In order to reconstruct the function of the object, In order to measure the objective function of the object, The scale weights representing the reconstructed objective function, Representing the scaling weight of the metrology target function.
  7. 7. The method for extracting rf characteristics of a radiation source based on metric and deep learning of claim 6, wherein the reconstruction objective function is: Wherein, the To model the processing results corresponding to the mth RF signal sample for the nth type of radiation source in the dataset, Is that And (3) outputting through the FRM network constrained by the output of the feature encoder of the AE network, wherein N is the total number of radiation source types, and M is the number of RF signal samples under each type of radiation source in the batch of samples.
  8. 8. The method for extracting RF characteristics from a radiation source based on metric and deep learning as claimed in claim 7, wherein said method comprises the steps of The output through the FRM network subject to AE network output constraints is: Wherein, the Reconstruction decoder for FRM network Is used for the control of the temperature of the liquid crystal display device, Is that And Is used for the splicing result of the (a), Is that Signal characteristic encoder of AE network after training Is provided with an output of (a), Is that RF feature encoder via FRM network An output of (2); Wherein, the Representing parameters of the RF signature encoder.
  9. 9. The method for extracting RF characteristics of a radiation source based on metric and deep learning as claimed in claim 8, wherein said metric objective function The calculation process of (1) is as follows: step 1), representing the processing results corresponding to the RF signal samples in the batch as After inputting the processing result corresponding to the RF signal samples in the batch of samples into the FRM network, the RF feature encoder of the FRM network The depth RF signature of the output is expressed as: ; Step 2), calculating depth RF feature centroid, and splicing the depth RF feature outputted in step 1) with the calculated depth RF feature centroid to form Pairs of RF characteristics to be composed of The RF feature pairs are used as inputs to a metric score; In the step 2), the depth RF feature output in the step 1) is spliced with the calculated depth RF feature centroid, which specifically includes: Wherein, the Representative of RF signal samples in a batch of samples The corresponding processing results are output through the RF signature encoder of the FRM network, Representative of when In the time-course of which the first and second contact surfaces, The centroid of the corresponding depth RF feature, Representative of when In the time-course of which the first and second contact surfaces, The centroid of the corresponding depth RF feature, Representative RF signal samples A splice result with a depth RF feature centroid; step 3), utilizing a measurement score device Calculating a similarity score between the RF features and the depth RF feature centroid of each RF feature pair: Wherein, the Parameters representing a metric score; Step 4) constructing a metric objective function based on the results of step 3) ; The specific process of the step 4) is as follows: Wherein, the As an intermediate variable, log is the logarithm based on e.
  10. 10. The method for extracting rf characteristics of a radiation source based on metric and deep learning as claimed in claim 9, wherein the following steps And The calculation process of (1) is as follows: Wherein, the Representing the output of the RF signature encoder of the FRM network from the processing result corresponding to the ith RF signal sample at the kth type of radiation source in the batch of samples, Representing and averaging; 。

Description

Radiation source radio frequency characteristic extraction method based on measurement and deep learning Technical Field The invention belongs to the technical field of radiation source radio frequency feature extraction, and particularly relates to a radiation source radio frequency feature extraction method based on measurement and deep learning. Background The radiation source Radio Frequency (RF) feature extraction technology is widely applied to various fields such as electronic countermeasure, spectrum management, cognitive radio, wireless network security and the like. The radiation source is the core of a system such as a radar and communication system that generates an RF signal of sufficient power and transmits it through an antenna. The radiation source has various non-ideal characteristics during production and operation, and the non-ideal characteristics of the radiation source can lead to deviation of an emitted Intentionally Modulated (IM) signal, so that the radiation source carries Unintentional Modulation (UM) containing hardware information, namely the radiation source is attached to the IM signal in the form of UM, and the UM is derived from a plurality of electronic components (such as a signal source, a mixer, a power amplifier and other modules) inside the radiation source and is expressed on the IM signal in the form of resultant force. Moreover, this non-ideal characteristic varies from device to device, is inherent to the radiation source, and is not counterfeitable and is difficult to change. Meanwhile, radiation sources with different UM parameters have different UM information (also called RF information), which may cause RF differences between RF signals. Different IM parameters having different IM information (also known as signal information) can result in signal differences between the RF signals. Thus, the RF signal is mainly affected by both signal information and RF information. Different UM parameters and IM parameters of the radiation source may generate RF signals carrying different RF information and signal information. How to more accurately characterize RF information, i.e. extract RF features, by means of received RF signals is a matter of great concern to current practitioners. A number of feature extraction methods have been proposed at present, and conventional feature extraction methods are mainly divided into a time domain, a frequency domain, a time-frequency domain, and a transform domain. The disadvantages of these methods are that manual feature extraction and design is dependent on expert knowledge and experience, and the methods and feature dimensions are high and feature levels are shallow. With reference to the successful combination of deep learning and radiation source identification, more and more students focus on how to automatically extract RF features using deep learning, but research is still in the start stage. The method is characterized in that the RF signals under constant UM parameters (namely, the same manufacturer and the same model of individual) and constant IM parameters (namely, the same modulation type and the same frequency and the like) are analyzed and subjected to feature extraction, namely, constant UM information and IM information, the assumption of the constant parameters is too ideal, and meanwhile, the main problems of the methods are that the RF signals are directly subjected to feature extraction, namely, the IM information and the UM information are prone to be analyzed as a whole, however, interference caused by the IM information is not concerned, and a large amount of IM information is contained in the extracted RF features. Meanwhile, with the rapid development of various technologies, RF signals emitted from radiation sources are more and more complex, and when different tasks are performed, RF signals with "random" variation of IM parameters, i.e., IM information, are generally adopted, and thus, IM information is increased. And as the circuit integration process improves, non-ideal characteristics between components are decreasing, i.e., UM information is weakening. In other words, weakening of the UM information implicitly increases the IM information, and increasing of the IM information implicitly weakens the UM information. The IM information is increased and the UM information is decreased, making it difficult for the conventional method to extract RF features with good stability and separability. In fact, UM carrying hardware-containing information occupies only a very small portion of the information conveyed on the RF signal, i.e., the signal information carried by the RF signal may cover the RF information, or may completely exceed the RF information. UM information is very weak and difficult to characterize compared to IM information. Since the RF signal contains both IM and UM information, but we are not concerned about the form of IM at this time, in the radiation source RF signature extraction techni