A method for extracting radio frequency features of a radiation source based on metrics and deep learning

Abstract

The application discloses a kind of based on metric and deep learning's radiation source radio frequency feature extraction method, it belongs to radiation source radio frequency feature extraction technical field.The present application solves the problem that it is difficult to extract RF feature with good stability and separability using existing method, and the extracted RF feature is easily disturbed by IM information and fails.The method of the present application is: constructing modeling dataset and ideal training dataset;Each RF signal sample in the constructed dataset is processed to obtain the processing result corresponding to each RF signal sample;The AE network built is trained using the processing result of the RF signal sample in the ideal training dataset;The FRM network is trained using the processing result of the RF signal sample in the modeling dataset and the trained AE network;After processing the RF signal to be detected, the processing result is input into the FRM network constrained by the signal feature encoder output by the trained, and the radiation source radio frequency feature extraction result is obtained.The present application can be applied to radiation source radio frequency feature extraction.

Description

Technical Field

[0001] This invention belongs to the field of radio frequency feature extraction technology for radiation sources, specifically relating to a method for extracting radio frequency features of radiation sources based on metric and deep learning. Background Technology

[0002] Radio frequency (RF) feature extraction technology for radiation sources has wide applications in many fields such as electronic countermeasures, spectrum management, cognitive radio, and wireless network security. Radiation sources are core components of systems such as radar and communications, generating a sufficiently powerful RF signal that is transmitted through an antenna. During production and operation, radiation sources exhibit various non-ideal characteristics. These inherent non-ideal characteristics cause deviations in the transmitted intentionally modulated (IM) signal, thus carrying unintentional modulation (UM) containing hardware information. This UM is added to the IM signal in the form of UM, originating from numerous electronic components within the radiation source (such as signal generators, mixers, power amplifiers, etc.), and manifesting as a "combined force" on the IM signal. Moreover, these non-ideal characteristics vary from device to device, are inherent to the radiation source, and are unforgeable and difficult to alter.

[0003] Meanwhile, radiation sources with different UM parameters possess different UM information (also known as RF information), leading to RF differences between RF signals. Similarly, different IM parameters possess different IM information (also known as signal information), resulting in signal differences between RF signals. Therefore, RF signals are primarily influenced by both signal information and RF information. Different UM and IM parameters of a radiation source can generate RF signals carrying different RF and signal information.

[0004] How to more accurately represent RF information from received RF signals, i.e., extract RF features, is a problem of great interest to scholars. Many feature extraction methods have been proposed, mainly categorized into time domain, frequency domain, time-frequency domain, and transform domain methods. The drawbacks of these methods are their reliance on expert knowledge and experience for manual feature extraction and design, and their relatively high method and feature dimensionality and shallow feature level. Drawing on the successful combination of deep learning technology and radiation source identification technology, more and more scholars are focusing on how to automatically extract RF features using deep learning technology, but research is still in its early stages.

[0005] Both traditional feature extraction methods and neural network-based feature extraction methods analyze and extract features from RF signals under constant UM parameters (i.e., individuals from the same manufacturer and model) and constant IM parameters (i.e., the same modulation type, the same frequency, etc.). This assumption of constant UM and IM information is too idealistic. At the same time, the main problem with these methods is that they directly extract features from RF signals, that is, they tend to analyze IM and UM information as a whole, without paying attention to the interference brought by IM information. As a result, the extracted RF features contain a large amount of IM information.

[0006] Meanwhile, with the rapid development of various technologies, the RF signals emitted by radiation sources are becoming increasingly complex. When performing different tasks, RF signals with "random" variations in IM parameters are typically used, meaning IM information is increased. Furthermore, with improvements in circuit integration technology, the non-ideal characteristics between components are decreasing, meaning UM information is weakening. In other words, the weakening of UM information implicitly increases IM information, and vice versa. This increase in IM information and decrease in UM information makes it difficult for traditional methods to extract RF features with good stability and separability. In fact, UM, carrying hardware information, accounts for only a very small portion of the information transmitted in the RF signal; that is, the signal information carried by the RF signal may cover or even completely exceed the RF information. Compared to IM information, UM information is very weak and difficult to characterize. Since the RF signal contains both IM and UM information, but we don't care about the form of IM, in radiation source RF feature extraction technology, IM information is not just useless information, but a powerful interference signal. This presents a serious challenge to radiation source RF feature extraction technology.

[0007] In summary, current research on radio frequency (RF) feature extraction techniques reveals that, under complex conditions, existing methods struggle to extract RF features with good stability and separability. The difficulty of RF feature extraction is significantly increased, and the lack of influence from variable parameter information (IM) on RF feature extraction makes traditional methods susceptible to interference from IM information, rendering the extracted RF features ineffective. Therefore, effectively reducing IM interference to "easily" extract RF features with good stability and separability is a pressing issue that needs to be addressed. Summary of the Invention

[0008] The purpose of this invention is to address the problem that existing methods are difficult to extract RF features with good stability and separability under complex conditions, and that the extracted RF features are easily invalidated by IM information interference. Therefore, this invention proposes a method for extracting radio frequency features of radiation sources based on metric and deep learning.

[0009] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0010] A method for extracting radio frequency features of radiation sources based on metric and deep learning, the method specifically includes the following steps:

[0011] Step 1: Model N types of radiation source structure models, and generate P radiation source individuals for each type of radiation source structure model by adjusting the UM parameter;

[0012] By adjusting the IM modulation parameters of each individual radiation source, an RF signal sample set carrying RF information is generated, and the generated RF signal sample set carrying RF information is used as the modeling dataset.

[0013] Based on the IM modulation parameters used when generating the modeling dataset, an RF signal sample set without carrying RF information is generated, and the generated RF signal sample set without carrying RF information is used as the ideal training dataset.

[0014] Step 2: Process each RF signal sample in the modeling dataset and the ideal training dataset respectively to obtain the processing result for each RF signal sample;

[0015] Step 3: Construct an AE network, which includes a signal feature encoder f. sig and reconstructed decoder f d The AE network is trained using the processing results of RF signal samples in the ideal training dataset.

[0016] Step 4: Design the FRM network, which includes an RF feature encoder g. rf 、Reconstructed decoder g d And metric scorer g s The FRM network is trained using the processing results of RF signal samples in the modeling dataset and the trained AE network.

[0017] Step 5: Acquire the RF signal to be detected, process the acquired signal, and input the processing result into the trained signal feature encoder f. sig The output-constrained FRM network will use the RF feature encoder g rf The output is used as the result of radio frequency feature extraction of the radiation source.

[0018] The beneficial effects of this invention are:

[0019] This invention first designs an AE network and an FRM network. The signal feature encoder trained in the AE network is used to fix the depth signal features, thereby constraining the RF feature encoder in the FRM network to automatically extract depth RF features that do not contain signal information. This reduces interference from signal information, making it less sensitive to IM information and giving the depth RF features good stability, while also establishing a one-to-one correspondence between RF information and RF characteristics. Secondly, a metric scorer is designed in the FRM network using a metric learning method. This reduces the distance between RF signal samples with different IM parameters from the same type of radiation source in the feature domain during the learning process, and increases the distance between RF signal samples from different types of radiation sources. The distance between samples in the feature domain ensures good separability of deep RF features. Finally, the RF feature encoder is optimized through unsupervised and supervised dual learning using the reconstruction decoder and metric scorer in the FRM network. This ensures that the deep RF features automatically extracted from non-ideal RF signals by the RF feature encoder possess both good stability and separability. This end-to-end training method eliminates inconsistencies in individual network training, leading to an optimal network. Furthermore, using the original time-domain signal as the network input avoids complex time-frequency transformations, preventing feature information loss and reducing preprocessing complexity. The input and objective functions of the FRM network are designed on a batch of samples, accelerating network training. Additionally, the absence of hierarchical initialization and fine-tuning simplifies the training process. This invention can "easily" extract RF features even in complex backgrounds, providing a new approach to RF feature extraction technology for radiation sources. Attached Figure Description

[0020] Figure 1 Links for different types of RES models;

[0021] Figure 2 RF signal spectrum diagrams for different IMs under different types of RES models;

[0022] Figure 3 Here is a flowchart of the deep RF feature extraction method;

[0023] Figure 4 Flowchart for generating RF signal dataset;

[0024] Figure 5 For AE network;

[0025] Figure 6 For FRM network;

[0026] Figure 7 For measuring scorers;

[0027] Figure 8Thermodynamic distribution maps of the depth RF characteristics of different RES;

[0028] Figure 9 A visualization of the depth RF features of CW signals under different RES values ​​in two-dimensional space;

[0029] Figure 10 A visualization of the depth RF features of BPSK signals under different RES values ​​in two-dimensional space;

[0030] Figure 11 A visualization of the depth RF features of LFM signals under different RES values ​​in two-dimensional space;

[0031] Figure 12 A visualization of the distribution of depth RF features of three IMs under different RES in two-dimensional space. Detailed Implementation

[0032] Specific Implementation Method 1: The radiation source radio frequency feature extraction method based on metric and deep learning described in this implementation method specifically includes the following steps:

[0033] Step 1: Model N types of radiation source structure models, and generate P radiation source individuals for each type of radiation source structure model by adjusting the UM parameter;

[0034] By adjusting the IM modulation parameters of each individual radiation source, a set of RF signal samples carrying RF information (non-ideal RF signals) is generated, and the generated set of RF signal samples carrying RF information is used as the modeling dataset.

[0035] Based on the IM modulation parameters used when generating the modeling dataset, an RF signal sample set without carrying RF information is generated, and the generated RF signal sample set without carrying RF information is used as the ideal training dataset.

[0036] Step 2: Process each RF signal sample in the modeling dataset and the ideal training dataset respectively to obtain the processing result for each RF signal sample;

[0037] Step 3: Construct an AE network, which includes a signal feature encoder f. sig and reconstructed decoder f d The AE network is trained using the processing results of RF signal samples in the ideal training dataset.

[0038] Step 4: Design the FRM network, which includes an RF feature encoder g. rf 、Reconstructed decoder g d And metric scorer g sThe FRM network is trained using the processing results of RF signal samples in the modeling dataset and the trained AE network.

[0039] Step 5: Acquire the RF signal to be detected, process the acquired signal, and input the processing result into the trained signal feature encoder f. sig The output-constrained FRM network will use the RF feature encoder g rf The output is used as the result of radio frequency feature extraction of the radiation source.

[0040] This embodiment designs and develops a radiation source radio frequency feature extraction method based on metric and deep learning. By combining metric learning and deep learning techniques, it studies how to effectively reduce interference from IM information to reduce the difficulty of radiation source RF feature extraction technology. This allows for the "easy" extraction of stable and separable deep RF features from the RF signals emitted by the radiation source under complex conditions, providing a completely new perspective on radiation source radio frequency feature extraction.

[0041] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that the RF signal sample that does not carry RF information is:

[0042] y id =x im =IM(s)

[0043] Among them, y id Here are RF signal samples that do not carry RF information, s is the baseband signal, and IM(·) is the IM system;

[0044] The RF signal sample carrying RF information is:

[0045] y nid =RF(x im =RF(IM(s))

[0046] Among them, y nid RF(·) represents an RF signal sample carrying RF information, and RF(·) represents an RF system.

[0047] The other steps and parameters are the same as in Specific Implementation Method 1.

[0048] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that the RF signal sample processing method is as follows:

[0049] For any RF signal sample, extract in-phase and quadrature data from the RF signal sample; then concatenate the in-phase and quadrature data.

[0050] x′=[x I ,x Q ]

[0051] Where, x I For in-phase data, x Q The data are orthogonal, and x′ is the concatenated result;

[0052] After performing maximum absolute value normalization on the concatenated result x′, the processing result corresponding to this RF signal sample is:

[0053]

[0054] Where x is the processing result corresponding to the RF signal sample, |x′ max | represents the maximum absolute value of the data in the concatenation result.

[0055] Other steps and parameters are the same as in specific implementation method one or two.

[0056] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that the optimizer used when training the constructed AE network is an adaptive moment estimation optimizer, the learning rate is set to 0.0001, and the objective function used is:

[0057]

[0058] Among them, L AE The objective function used when training the AE network, x i This represents the processing result corresponding to the i-th RF signal sample in the ideal training dataset. For x i The output of the AE network is ||·||, which is the 2-norm, and Q represents the batch size.

[0059] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0060] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that the x... i The output after passing through the AE network is:

[0061] h′ sig =f sig (x i ,ω sig )

[0062]

[0063] Among them, f sig For signal feature encoder, ω sig h′ is a parameter of the signal feature encoder. sig f is the output of the signal feature encoder. d For reconstructing the decoder, ω d These are the parameters for reconstructing the decoder.

[0064] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0065] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One through Five in that the optimizer used when training the FRM network is an adaptive moment estimation optimizer, the learning rate is set to 0.0001, and the objective function used is:

[0066] L FRM =λ1L r +λ2L MS

[0067] Among them, L FRM The objective function used when training the FRM network, L r To reconstruct the objective function, L MS To measure the objective function, λ1 represents the proportional weight of the reconstruction objective function, and λ2 represents the proportional weight of the measurement objective function.

[0068] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0069] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that the reconstruction objective function is:

[0070]

[0071] Among them, y nm To model the processing results corresponding to the m-th RF signal sample under the n-th type of radiation source in the dataset, For y nm The output of the FRM network, constrained by the feature encoder output of the AE network, is given by N, which represents the total number of radiation source classes, and M, which represents the number of RF signal samples for each radiation source class in the batch.

[0072] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0073] Specific Implementation Method Eight: This implementation method differs from one of Specific Implementation Methods One to Seven in that the y nm The output of the FRM network, constrained by the AE network output, is:

[0074]

[0075] Where, θ d For the reconstruction decoder g of the FRM network d The parameter h rs for h rf(nm) and h sig The splicing result, h sig For ynm The signal feature encoder f of the trained AE network sig The output, h rf(nm) For y nm RF feature encoder g after FRM network rf The output;

[0076] h rf(nm) =g rf (y nm ,θ rf )

[0077] Where, θ rf This represents the parameters of the RF feature encoder.

[0078] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0079] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that the objective function L is described in the embodiment. MS The calculation process is as follows:

[0080] Step 1) Represent the processing results corresponding to the RF signal samples in the batch as [(y 11 ,y 12 ,…,y 1M ),(y 21 ,y 22 ,…,y 2M ),…,(y N1 ,y N2 ,…,y NM After inputting the processing results corresponding to the RF signal samples in the batch into the FRM network, the RF feature encoder g of the FRM network is then used. rf The output depth RF features are represented as: [(h rf(11) ,h rf(12) ,…,h rf(1M) ),(h rf(21) ,h rf(22) ,…,h rf(2M) ),…,(h rf(N1) ,h rf(N2) ,…,h rf(NM) )];

[0081] Step 2) Calculate the centroid of the deep RF feature. Concatenate the deep RF feature output in Step 1) with the calculated centroid of the deep RF feature to form N×M×N RF feature pairs. Use the N×M×N RF feature pairs as the input of the metric scorer.

[0082] In step 2), the depth RF features output in step 1) are concatenated with the calculated depth RF feature centroids, specifically as follows:

[0083]

[0084] Among them, h rf(nm) y represents the RF signal sample in the batch. nm The corresponding processing result is output by the RF feature encoder of the FRM network. This represents h when k = n. rf(nm) The corresponding depth RF feature centroid, c k This represents h when k≠n. rf(nm) The corresponding depth RF feature centroid, h rf(nmk) Represents RF signal sample y nm The result of splicing with the centroids of deep RF features;

[0085] Step 3) Using the metric scorer g s Calculate the similarity score between the RF features and the centroids of the deep RF features in each RF feature pair:

[0086] s nmk =g s (h rf(nmk) ,θ s )

[0087] Where, θ s This represents the parameters of the scorer.

[0088] Step 4) Construct the objective function L based on the results of Step 3). MS ;

[0089] The specific process of step 4) is as follows:

[0090]

[0091]

[0092] Among them, L ms As an intermediate variable, log is the logarithm to the base e.

[0093] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0094] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One to Nine in that the c... k and c( k The calculation process for -m) is as follows:

[0095]

[0096] Among them, h rf(ki)The output of the RF feature encoder of the FRM network represents the processing result of the i-th RF signal sample under the k-th radiation source in the batch. i [·] indicates calculating the mean;

[0097]

[0098] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0099] Example

[0100] The present invention will now be described in further detail with reference to the accompanying drawings, such as... Figures 1 to 7 As shown, the radio frequency feature extraction method for radiation sources provided by the present invention includes the following steps:

[0101] 1. Modeling the Radiation Source: Based on the working principle of the radiation source, a DDS signal source model and three types of RF link models (PA1, PA2, and PA3) were built using a simulation platform with measurement parameters by selecting different UM parameters. By combining the DDS signal source model and the RF link model, three types of radar radiation source structural models (RES1, RES2, and RES3) were constructed, as follows... Figure 1 As shown. In addition, under the same RES model, four radiation source individual (REI) models were built by slightly adjusting the UM parameters of each module, so that there are structural differences between different types of RES models, and individual differences between different REI models under the same RES model.

[0102] 2. Dataset Generation: Based on the above modeling, firstly, by adjusting different IM modulation parameters (including modulation type MT, frequency F, bandwidth B, and input power P), source signals with non-ideal characteristics of three modulation types are generated: continuous wave (CW), linear frequency modulation (LFM), and binary phase shift keying (BPSK). These are then amplified by different types of RF link models to generate non-ideal RF signals under different parameters, thus constructing a modeling dataset in a complex context. Secondly, 90% of the samples are divided into a modeling training dataset, and the remaining 10% is used as a modeling test dataset. Figure 2 As shown, the spectra of RF signals generated by the three types of RES models are clearly different in terms of spectral purity under different modulation types. Finally, an ideal RF signal without RF information is generated based on the IM parameters of the modeling dataset to construct the corresponding ideal training dataset.

[0103] 3. RF signal model:

[0104] We believe that for IM signal x im It can be considered that the baseband signal s is generated by the IM system IM(·), that is:

[0105] x im =IM(s)

[0106] A radiation source can also be considered as an RF system RF(·) carrying RF information. If the RF system is ideal, that is, the RF signal emitted by the radiation source does not carry RF information, then the ideal RF signal model can be expressed as:

[0107] y id =x im

[0108] If the RF system is non-ideal, that is, the RF signal model transmitted by the RF system carrying RF information can be represented as:

[0109] y nid =RF(x im =RF(IM(s))

[0110] In fact, an RF system is a complex, nonlinear, and unknown system; therefore, its specific parameters cannot be obtained through simple calculations. Furthermore, different types of radiation sources carry different RF information and correspond to different RF systems. Therefore, different UM and IM parameters can generate various RF signals carrying different RF and signal information, i.e., a mapping from the parameter domain to the data domain.

[0111] Therefore, the idea behind this invention is: if the data domain can be mapped to the feature domain, enabling the neural network to automatically learn relevant features representing RF information and signal information respectively, and using signal features to constrain RF features in the feature domain, then RF features in RF signals with variable IM parameters can be extracted "easily." To distinguish this from RF features manually extracted using traditional methods, this invention renames the RF features automatically extracted using deep learning as "deep RF features," such as... Figure 3 The diagram shown is a flowchart of the deep RF feature extraction method.

[0112] Furthermore, it should be noted that this invention performs RF feature extraction at the radiation source structure level, which is fundamentally different from RF feature extraction at the individual radiation source level. The individual level refers to the differences between different individual radiation sources, while the structural level focuses on the differences between the structures of different types of radiation sources. That is, in the radiation source modeling of this invention, such as... Figure 4 As shown, a similar radiation source structure can contain many different individual radiation sources. In this invention, these are considered similar radiation source structures, and the extracted features are similar. This technique of studying the RF features of radiation sources at a complex level is challenging. Specifically, the complexity in this paper is manifested in two aspects: firstly, the complexity of the radiation source types, i.e., the complexity of the UM parameters; and secondly, the complexity and variability of the IM parameters.

[0113] 4. Data Preprocessing: Preprocess the RF signals from the modeling dataset and the ideal training dataset. First, extract in-phase / quadrature (I / Q) data from the RF signals, then concatenate them and perform maximum absolute value normalization. The final RF signal x is then obtained. n Represented as:

[0114] x′=[x I ,x Q ]

[0115]

[0116] 5. Building the AE network: The AE network consists of a signal feature encoder f sig and a reconstruction decoder f d Composition, such as Figure 5 As shown, a signal feature encoder is used to automatically learn the depth signal features h in an RF signal. sig Building an AE network involves the input signals of the signal feature encoder and the reconstruction decoder, the network structure, and the objective function.

[0117] 1) Signal feature encoder:

[0118] Input signal: The input of the signal feature encoder is an ideal RF signal x that does not carry RF information, and the number of input nodes of the network is set to 1000×2=2000.

[0119] Network Structure: The signal feature encoder is configured with two hidden layers, where the number of nodes in the hidden layers is less than the number of nodes in the input layers. The number of output nodes is set to 50, meaning the depth signal features are 50-dimensional. The number of nodes in each layer is set to 2000-600-150-50. The encoding process of the network can be described as inputting an ideal RF signal into the signal feature encoder to automatically learn the depth signal features, achieving a mapping from the data domain to the feature domain, i.e.:

[0120] h s ′ ig =f sig (x i ,ω sig )

[0121] Where, ω sig These represent the parameters of the signal characteristic encoder.

[0122] 2) Reconstructing the decoder:

[0123] Input signal: depth signal features h automatically extracted by the signal feature encoder sig As input to the reconstruction decoder in the AE network, the number of input nodes in the network is 50.

[0124] Network Structure: The reconstruction decoder is configured with two hidden layers, and the number of nodes in each layer is set to 50-150-600-2000, meaning the network output has 2000 nodes. The decoding process can be described as using unsupervised learning to input the depth signal features automatically extracted by the signal feature encoder into the reconstruction decoder to reconstruct the RF signal. To achieve the mapping from the feature domain to the data domain.

[0125]

[0126] Where, ω d This represents the parameters of the reconstruction decoder.

[0127] Objective function: Utilizing unsupervised learning to achieve a mapping from the data domain to the feature domain and back to the data domain. The mean squared error is used to establish the objective function of the AE network, expressed as:

[0128]

[0129] 6. Training the AE network:

[0130] The ideal training dataset is preprocessed and then used as input to the AE network, using the objective function L. AE The AE network was trained using the Adaptive Moment Estimation (Adam) optimizer with a learning rate of 0.0001 and a batch size of 60. The trained network was then saved, and the parameters ω of the signal feature encoder and reconstruction decoder were obtained. sig ,ω d .

[0131] 7. Design the FRM network:

[0132] The FRM network consists of an RF feature encoder g rf (Feature Encoder), Reconstruction Decoder g d (ReconstructionDecoder) and metric scorer g s (Metric Scorer) consists of, such as Figure 6 As shown, the signal feature encoder trained in the AE network is used to fix the depth signal features. Combined with the reconstruction decoder in the FRM network, the RF feature encoder in the FRM network can automatically learn the depth RF features h that do not contain signal information by constraining the feature domain. rfThe FRM network maps RF information to RF features one-to-one. It then utilizes the reconstruction decoder and metric scorer within the FRM network to optimize the RF feature encoder through both unsupervised and supervised learning. This ensures that the deep RF features automatically extracted from non-ideal RF signals by the RF feature encoder exhibit both stability and good separability. The design of the FRM network includes the input signals, network structure, and objective function of the RF feature encoder, reconstruction decoder, and metric scorer.

[0133] 1) RF Feature Encoder

[0134] Input signal: The input of the RF feature encoder is a non-ideal RF signal y carrying RF information, and the number of input nodes of the network is set to 1000×2=2000.

[0135] Network Structure: The RF feature encoder is configured with two hidden layers, where the number of nodes in the hidden layers is less than the number of nodes in the input layers. The number of output nodes is set to 50, meaning the deep RF features are 50-dimensional. The number of nodes in each layer is set to 2000-600-150-50. The encoding process of the network can be described as inputting a non-ideal RF signal into the RF feature encoder to automatically learn deep RF features, achieving a mapping from the data domain to the feature domain.

[0136] h rf(nm) =g rf (y nm ,θ rf )

[0137] Where, θ rf This represents the parameters of the RF feature encoder.

[0138] 2) Reconstructing the decoder

[0139] Input signals: Non-ideal RF signals are input into the RF feature encoder in the FRM network and the signal feature encoder in the trained AE network, respectively, to automatically learn deep RF features and extract deep signal features. The deep RF features and deep signal features are then concatenated and used as input to the reconstruction decoder. Therefore, the number of input nodes in the network is 50 × 2 = 100.

[0140] h rs =[h rf ,h sig ]

[0141] Network Structure: The reconstruction decoder is configured with two hidden layers, and the number of nodes in each layer is set to 50-150-600-2000, meaning the network output has 2000 nodes. The decoding process can be described as using unsupervised learning to input the deep RF features automatically extracted by the RF feature encoder into the reconstruction decoder to reconstruct the RF signal. To achieve the mapping from the feature domain to the data domain.

[0142]

[0143] Where, θ d This represents the parameters of the reconstruction decoder.

[0144] Objective function: To achieve the mapping from the data domain to the feature domain and back to the data domain using unsupervised learning. The objective function for reconstructing the FRM network using mean squared error is established as follows:

[0145]

[0146] 3) Metric scorer

[0147] Metric learning is one of the core problems in pattern recognition. Metric learning can also be considered as similarity measurement to measure the similarity between signal samples. The purpose of using metric learning in this invention is to enable the network to reduce the distance between RF signal samples with different IM parameters from the same type of radiation source in the feature domain, and to increase the distance between RF signal samples from different types of radiation sources in the feature domain during the learning process. Therefore, the metric learning method aims to improve the separability of deep RF features automatically extracted by the RF feature encoder in the feature domain.

[0148] Input signal: Based on the design concept of the generalized end-to-end objective function, the input of the metric scorer is set to a batch of samples, including N different types of radiation sources, and M samples under each type of radiation source, such as... Figure 7 As shown. Therefore, firstly, it is necessary to acquire (N×M) RF signals y. nm To construct a batch sample, we use (1≤n≤N, 1≤m≤M), i.e.:

[0149] [(y 11 ,y 12 ,…,y 1M ),(y 21 ,y 22 ,…,y 2M ),…,(y N1 ,y N2 ,…,y NM )]

[0150] Then, it is input into the RF feature encoder to automatically extract (N×M) deep RF features h of this batch of samples. rf(nm) (representing the m-th depth RF feature of the n-th type of radiation source), that is:

[0151] [(h rf(11) ,h rf(12) ,…,h rf(1M) ),(h rf(21) ,hrf(22) ,…,h rf(2M) ),…,(h rf(N1) ,h rf(N2) ,…,h rf(NM) )]

[0152] Secondly, the depth RF features h of the M samples of the k-th type of radiation source are calculated respectively. rf(ki) =(h rf(k1) ,h rf(k2) ,…,h rf(kM) The depth RF eigencentroids of (k=1,…,N,i=1,…,M) are:

[0153]

[0154] Finally, the (N×M) deep RF features in a batch of samples are concatenated with the centroids of N deep RF features to form (N×M×N) RF feature pairs, which are used as inputs to the metric scorer.

[0155] When k = n, removing the deep RF features of this sample when calculating the centroid can make the network training more stable. Therefore, when k = n, the deep RF feature centroid is calculated using the following formula:

[0156]

[0157] Therefore, (N×M×N) RF feature pairs are represented as:

[0158]

[0159] Network Structure: The metric scorer is configured with two hidden layers, and the number of nodes in each layer is set to 100-50-20-1, meaning the network has 100 input nodes and 1 output node. The network's metric scoring process can be described as inputting RF feature pairs into the metric scorer to calculate the similarity between the deep RF features and the deep RF feature centroids.

[0160] The similarity score between each RF feature pair is calculated using a metric scorer, and can be represented as:

[0161] s nmk =g s (h rf(nmk) ,θ s )

[0162] Where, θ s This represents the parameters of the scorer.

[0163] Objective function: In the metric scorer, for each of the M samples of each type of radiation source, each deep RF feature should have higher similarity to the centroid of deep RF features of the same type of radiation source and be far away from the centroid of deep RF features of different types of radiation sources. That is, when k = n, it indicates that the deep RF features and their centroids in the RF feature pair belong to the same type, i.e., positive samples; when k ≠ n, it indicates that the deep RF features and their centroids in the RF feature pair belong to different types, i.e., negative samples. For example... Figure 7 As shown, the expected effect of the metric scorer is to give larger values ​​to the colored parts of the similarity score and smaller values ​​to the gray parts.

[0164] When calculating the objective function, the (N×M×N) similarity scores of a batch of samples are first mapped to a similarity score matrix of dimension (N×M,N), that is:

[0165] Map the similarity scores calculated in step 3) to a similarity matrix s of dimension (N×M,N):

[0166]

[0167] Where the superscript T represents the transpose of the matrix;

[0168] To map the similarity score range to 0-1, the SoftMax function is applied to the similarity score of each row of the matrix, and then the similarity score is regressed to the cross-entropy (CE) objective function. When k=n, i.e., for positive samples, the output of the scorer is equal to 1; otherwise, for negative samples, the output is equal to 0. Therefore, the objective function for measuring the similarity score of a sample is expressed as:

[0169]

[0170] The first term is the objective function for positive samples, and the second term is the objective function for negative samples.

[0171] Therefore, the objective function for measuring a batch of samples is expressed as:

[0172]

[0173] Overall objective function:

[0174] In the FRM network, the RF feature encoder, reconstruction decoder, and metric scorer are treated as a single, integrated neural network. These three sub-networks are jointly trained end-to-end to optimize them in one direction. This eliminates inconsistencies in the training of individual networks, leading to an optimal network. Therefore, the overall objective function of the FRM network is expressed as:

[0175] L FRM =λ1Lr +λ2L MS

[0176] Where λ1 and λ2 represent the proportional weights of the reconstruction objective function and the measurement objective function, respectively.

[0177] 8. Training the FRM network:

[0178] The preprocessed training data is used as the input to the FRM network, and the objective function L is applied. FRM The FRM network was trained using the Adam optimizer with a learning rate of 0.0001 and a batch size of 3×20, meaning each batch contained 20 samples from three different radiation sources. The trained network was then saved, and the parameters θ of the RF feature encoder, reconstruction decoder, and metric scorer were obtained. rf ,θ d ,θ s .

[0179] 9. Construct a deep RF feature centroid database: In the last network training, save the deep RF feature centroids automatically extracted by the RF feature encoder for each type of radiation source's emitted RF signal calculated in each batch, and average them to construct a deep RF feature centroid database.

[0180] 10. Test the FRM network:

[0181] The FRM network was tested using a modeling test dataset. The test data was preprocessed and used as input to the FRM network. Deep RF features were automatically extracted using a pre-trained RF feature encoder, followed by separability and stability analysis.

[0182] The simulation results of the method of this invention are as follows:

[0183] Figure 8 The image shows the heatmap of 50-dimensional deep RF features extracted from three types of RES models by the FRM network. It can be seen that there are significant differences in the deep RF features extracted from the three types of RES models by the FRM network, proving that RF information and RF characteristics are mapped one-to-one, and RF feature extraction is performed at the structural level.

[0184] Figure 9 , Figure 10 , Figure 11 The image shows the distribution of deep RF features extracted by the FRM network under three different IMs (CW, BPSK, LFM) in two-dimensional space. It can be seen that the deep RF features extracted by the three RES models under different IMs all have good distinguishability.

[0185] Figure 12The image shows the distribution of deep RF features extracted by the FRM network under three hybrid IMs (CW, BPSK, and LFM) in two-dimensional space. It can be seen that the deep RF features extracted by the three RES models under different IMs have good discriminability, and different IMs under the same RES model are clustered together, which verifies that the deep RF features under variable parameters have good stability.

[0186] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for extracting radio frequency features of radiation sources based on metric and deep learning, characterized in that, The method specifically includes the following steps: Step 1: Model N types of radiation source structure models. By adjusting the UM parameter, generate P radiation source individuals for each type of radiation source structure model. UM originates from the electronic components inside the radiation source and is attached to the IM signal in the form of UM. By adjusting the IM modulation parameters of each individual radiation source, an RF signal sample set carrying RF information is generated, and the generated RF signal sample set carrying RF information is used as the modeling dataset. Based on the IM modulation parameters used when generating the modeling dataset, an RF signal sample set without carrying RF information is generated, and the generated RF signal sample set without carrying RF information is used as the ideal training dataset. Step 2: Process each RF signal sample in the modeling dataset and the ideal training dataset respectively to obtain the processing result for each RF signal sample; Step 3: Construct an AE network, which includes a signal feature encoder. and reconstructed decoder The AE network is trained using the processing results of RF signal samples in the ideal training dataset. Step 4: Design the FRM network, which includes an RF feature encoder. , reconstructed decoder and metric scorer The FRM network is trained using the processing results of RF signal samples in the modeling dataset and the trained AE network. The input to the RF feature encoder is an RF signal carrying RF information; The input to the reconstruction decoder is the concatenation result of deep RF features and deep signal features. Specifically, the RF signal carrying RF information is input into the RF feature encoder in the FRM network and the signal feature encoder in the trained AE network, respectively, so as to automatically learn deep RF features and extract deep signal features. The metric scorer calculates a similarity score between each pair of RF features; Step 5: Acquire the RF signal to be detected, process the acquired signal, and input the processing result into the trained signal feature encoder. The output-constrained FRM network will be used for the RF feature encoder. The output is used as the result of radio frequency feature extraction of the radiation source.

2. The method for extracting radio frequency features of radiation sources based on metric and deep learning according to claim 1, characterized in that, The RF signal sample that does not carry RF information is: in, An RF signal sample that does not carry RF information. For baseband signals, For IM system; The RF signal sample carrying RF information is: in, An RF signal sample carrying RF information. It is an RF system.

3. The method for extracting radio frequency features of radiation sources based on metric and deep learning according to claim 2, characterized in that, The RF signal sample is processed as follows: For any RF signal sample, extract in-phase and quadrature data from the RF signal sample; then concatenate the in-phase and quadrature data. in, For in-phase data, For orthogonal data, This is the result of a series connection. For the cascaded results After performing maximum absolute value normalization, the processing result for this RF signal sample is: in, This is the processing result corresponding to the RF signal sample. This represents the maximum absolute value of the data in the concatenated result.

4. The method for extracting radio frequency features of radiation sources based on metric and deep learning according to claim 3, characterized in that, The optimizer used when training the constructed AE network is an adaptive moment estimation optimizer, with a learning rate set to 0.0001, and the objective function is: in, The objective function used when training the AE network. For the first in the ideal training dataset Processing results corresponding to each RF signal sample for After the output of the AE network It is a 2-norm. This represents the batch size.

5. The method for extracting radio frequency features of radiation sources based on metric and deep learning according to claim 4, characterized in that, The The output after passing through the AE network is: in, For signal feature encoders, These are the parameters of the signal feature encoder. The output of the signal feature encoder, To reconstruct the decoder, These are the parameters for reconstructing the decoder.

6. The method for extracting radio frequency features of radiation sources based on metric and deep learning according to claim 5, characterized in that, The optimizer used when training the FRM network is an adaptive moment estimator, with a learning rate set to 0.0001, and the objective function is: in, The objective function used when training the FRM network. To reconstruct the objective function, To measure the objective function, The proportional weights represent the reconstructed objective function. This represents the proportional weight of the objective function.

7. The method for extracting radio frequency features of radiation sources based on metric and deep learning according to claim 6, characterized in that, The reconstruction objective function is: in, To model the processing results corresponding to the m-th RF signal sample under the n-th type of radiation source in the dataset, for The output of the FRM network, constrained by the feature encoder output of the AE network, is given by N, which represents the total number of radiation source classes, and M, which represents the number of RF signal samples for each radiation source class in the batch.

8. The method for extracting radio frequency features of radiation sources based on metric and deep learning according to claim 7, characterized in that, The The output of the FRM network, constrained by the AE network output, is: in, Reconstruction decoder for FRM network The parameters, for and The splicing result, for The signal feature encoder of the trained AE network The output, for RF feature encoder via FRM network The output; in, This represents the parameters of the RF feature encoder.

9. The method for extracting radio frequency features of radiation sources based on metric and deep learning according to claim 8, characterized in that, The objective function of measurement The calculation process is as follows: Step 1) Represent the processing results corresponding to the RF signal samples in the batch as follows: After inputting the processing results corresponding to the RF signal samples in the batch into the FRM network, the RF feature encoder of the FRM network is then... The output depth RF features are represented as follows: ; Step 2) Calculate the centroid of the depth RF feature, and concatenate the depth RF feature output in Step 1) with the calculated depth RF feature centroid to form the core. Each RF feature pair will be composed of Each RF feature pair is used as input to the metric scorer; In step 2), the depth RF features output in step 1) are concatenated with the calculated depth RF feature centroids, specifically as follows: in, RF signal samples representing a batch of samples The corresponding processing result is output by the RF feature encoder of the FRM network. Representative hour, The corresponding depth RF feature centroid, Representative hour, The corresponding depth RF feature centroid, Representative RF signal sample The result of splicing with the centroids of deep RF features; Step 3) Using a scorer Calculate the similarity score between the RF features and the centroids of the deep RF features in each RF feature pair: in, This represents the parameters of the scorer. Step 4) Construct the objective function based on the results of Step 3). ; The specific process of step 4) is as follows: in, As an intermediate variable, log is the logarithm to the base e.

10. The method for extracting radio frequency features of radiation sources based on metric and deep learning according to claim 9, characterized in that, The and The calculation process is as follows: in, The output of the RF feature encoder of the FRM network represents the processing result of the i-th RF signal sample under the k-th radiation source in the batch. This indicates calculating the mean. 。

Images