Method and system for specific emitter identification based on multi-sequence feature learning

Abstract

The present application relates to the technical field of radiation source identification, and particularly relates to a specific radiation source identification method and system based on multi-sequence feature learning, which generates simulated signals of I / Q path in-phase and quadrature signals of multiple modulation types according to the mechanism of radiation source fingerprint characteristics, and uses the simulated signals of the radiation source to build a sample data set; the signals in the sample data set are standardized to obtain multi-sequence signals; a sequence fusion convolution network model is constructed and trained and optimized using the multi-sequence signals in the sample data set; the radiation source signal to be identified is input into the trained and optimized sequence fusion convolution network model, and the modulation category of the radiation source signal to be identified is obtained using the trained and optimized sequence fusion convolution network model. The present application can improve the efficiency of radiation source identification, can identify multiple types of illegal entry radio stations, can ensure the security of communication networks, and can provide a basis for counter-decision in the field of counter-reconnaissance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radiation source identification, and particularly relates to a specific emitter identification method and system based on multi-sequence feature learning. BACKGROUND

[0002] Specific Emitter Identification (SEI) is a method of realizing the unique identification of a target emitter by extracting Radio Frequency Fingerprint (RFF) features that can reflect the individual differences of emitters. In the civil field, SEI can be used to identify illegal radio stations entering the country, thereby ensuring the security of the communication network and having certain academic research value. SEI methods can be broadly divided into two categories: traditional SEI methods and SEI methods based on deep learning. Traditional SEI methods can be divided into three categories according to the types of extracted features: signal parameter statistical features, signal transform domain statistical features, and mechanism model features. These three types of features have achieved fruitful results in different stages and scenarios, but the common problem is that they are limited by human understanding of the essence of signals, mathematical tools, and the mechanism of emitters, and it is difficult to understand and extract complex essential features, which makes it difficult to meet the requirements of universality, stability, and comprehensiveness of SEI.

[0003] In recent years, deep learning has been widely used in the field of SEI. Vector diagrams, variational modal decomposition waveforms, multi-transform domain features, and multi-projection features are used as original features for RFF extraction by neural networks, and have achieved good results. However, there are three problems with SEI methods based on deep learning: first, some SEI methods use single features, and the recognition performance of recognition methods using only one type of transform domain information varies in different scenarios, different channels, and noise conditions; second, although some SEI methods use multiple features, they do not design effective and reliable feature extraction networks for multiple feature inputs, resulting in suboptimal feature fusion and RFF information extraction; third, most SEI methods perform mathematical operations on the original signal waveform to convert it into other graphical features or sequence features before extracting the fingerprint features. With the mathematical operation process of the signal, the RFF features of the signal will inevitably be lost during the operation process, making it difficult for deep networks to fully utilize the RFF features contained in the signal. SUMMARY

[0004] Therefore, the present application provides a specific emitter identification method and system based on multi-sequence feature learning, which solves the problems of single feature, poor feature fusion network design, and ignoring sequence features in existing emitter individual identification methods, thereby improving the recognition rate.

[0005] According to the design scheme provided by the application, a specific radiation source identification method based on multi-sequence feature learning is provided, comprising:

[0006] According to the radiation source fingerprint feature generation mechanism, I / Q in-phase and quadrature signal radiation source simulation signals of multiple modulation types are generated, and the radiation source simulation signals are used to build a sample data set;

[0007] The signals in the sample data set are standardized to obtain a multi-sequence signal composed of simulation signal I / Q sequences, amplitude and phase A / P sequences, and frequency domain A / P sequences;

[0008] A sequence fusion convolutional network model for radiation source classification and identification is constructed, and the multi-sequence signals in the sample data set are used to train and optimize the sequence fusion convolutional network model;

[0009] The radiation source signal to be identified is input into the trained and optimized sequence fusion convolutional network model, and the trained and optimized sequence fusion convolutional network model is used to obtain the modulation category of the radiation source signal to be identified.

[0010] As the specific radiation source identification method based on multi-sequence feature learning of the application, further, according to the radiation source fingerprint feature generation mechanism, radiation source simulation signals of multiple modulation types are generated, comprising:

[0011] First, according to the radiation source signal fingerprint feature generation mechanism, and combining the distortion performance of the quadrature modulator, filter, oscillator and power amplifier, a transmitter distortion model is constructed;

[0012] Then, according to the transmitter distortion model, the distortion model parameters are adjusted to generate radiation source simulation signals.

[0013] As the specific radiation source identification method based on multi-sequence feature learning of the application, further, the radiation source simulation signals are used to build a sample data set, comprising: sampling each radiation source signal information point in the simulation signal, grouping n information points collected continuously each time to form a signal sample, and collecting m signal samples for each modulation type of radiation source simulation signal; according to the signal samples of all types of radiation source simulation signals, a sample data set is formed, and according to a preset proportion, the sample data set under each modulation type is divided to obtain a training sample set, a verification sample set and a test sample set for network model training and optimization, wherein n and m are preset thresholds.

[0014] As the specific radiation source identification method based on multi-sequence feature learning of the application, further, the signals in the sample data set are standardized, comprising:

[0015] Firstly, each in-phase quadrature signal in the signal sample is normalized to an I / Q sequence, an amplitude / phase A / P sequence and a frequency domain A / P sequence;

[0016] Then, the normalized signal is sampled to obtain a baseband signal complex sequence with a length of N, and the instantaneous amplitude and the instantaneous phase of the signal are obtained according to the baseband signal complex sequence;

[0017] Then, the baseband signal complex sequence is subjected to Fourier transform to obtain a frequency domain complex sequence representation, and the amplitude spectrum and the phase spectrum of the signal are obtained according to the frequency domain complex sequence representation;

[0018] Finally, each in-phase quadrature signal in the signal sample and the corresponding instantaneous amplitude, instantaneous phase, amplitude spectrum and phase spectrum are arranged to obtain a multi-sequence signal of the signal sample.

[0019] As the specific emitter identification method based on multi-sequence feature learning of the present application, further, the multi-sequence signal representation is: Wherein, r I (), r Q (), A(), P(), F(), respectively represent the in-phase component in the baseband signal complex sequence, the quadrature component in the baseband signal complex sequence, the instantaneous amplitude, the instantaneous phase, the amplitude spectrum and the phase spectrum, and N is the length of the baseband signal complex sequence.

[0020] As the specific emitter identification method based on multi-sequence feature learning of the present application, further, the sequence fusion convolutional network model for emitter classification and identification comprises: a feature fusion module for multi-sequence feature fusion, a squeeze-and-excitation module for recalibrating a sequence feature channel weight vector according to the sharing degree of features in an input feature vector to generate a new feature vector, and a time sequence convolutional classification module for classifying and identifying the new feature sequence.

[0021] As the specific emitter identification method based on multi-sequence feature learning of the present application, further, the feature fusion module utilizes three multi-channel convolutional blocks to respectively perform parallel convolution and splicing on the I / Q sequence, the time domain A / P sequence and the frequency domain A / P sequence of the I / Q dual-channel sequence to form a multi-sequence signal feature.

[0022] As the specific emitter identification method based on multi-sequence feature learning of the present application, further, the process of recalibrating the sequence feature channel weight vector according to the sharing degree of features in the input feature vector to generate the new feature vector in the squeeze-and-excitation module is represented as: Wherein, is the i-th element of the new feature vector . iThe channel weight of the i-th element X of the channel weight vector S of the input feature vector X is i The channel weight of the i-th element X of the channel weight vector S of the input feature vector X is

[0023] As the specific radiation source identification method based on multi-sequence feature learning of the present application, further, the time sequence convolution classification module classifies and identifies new feature sequences, comprising:

[0024] First, a plurality of time sequence residual blocks connected in turn at the input and output ends are used to extract information in the new feature sequence, wherein each time sequence residual block extracts time information of the input feature and obtains the output feature based on the input feature and the extracted time information;

[0025] Then, the output feature of the last time sequence residual block is classified by using the activation function Softmax of the full connection layer to obtain the type of the radiation source signal.

[0026] Further, the present application also provides a specific radiation source identification system based on multi-sequence feature learning, comprising a data simulation module, a data processing module, a model construction module and a target identification module, wherein,

[0027] The data simulation module is used to generate I / Q signal radiation source simulation signals of multiple modulation types according to the generation mechanism of the radiation source fingerprint features, and the sample data set is constructed by using the radiation source simulation signals;

[0028] The data processing module is used to standardize the signals in the sample data set to obtain a multi-sequence signal composed of analog simulation signal I / Q sequences, amplitude and phase A / P sequences and frequency domain A / P sequences;

[0029] The model construction module is used to construct a sequence fusion convolution network model for radiation source classification and identification, and the sequence fusion convolution network model is trained and optimized by using the multi-sequence signals in the sample data set;

[0030] The target identification module is used to input the radiation source signal to be identified into the trained and optimized sequence fusion convolution network model, and the modulation category of the radiation source signal to be identified is obtained by using the trained and optimized sequence fusion convolution network model.

[0031] The present application has the following beneficial effects:

[0032] The application extracts multiple sequence signals of a communication radiation source transmitting signal, and combines the sequence signals to form a multi-sequence feature, avoids loss of radio frequency fingerprint information in the process of transforming into other transform domains, fully fuses spatial features and timing features in the multi-sequence features, enhances the feature extraction capability and recognition efficiency of the network through a channel attention mechanism, can identify multiple radiation source types of illegal entry radio stations, ensures the security of the communication network, and can also provide a basis for counter-decision in the field of counter-reconnaissance. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 A specific radiation source identification principle based on multi-sequence feature learning in the embodiment is shown;

[0034] Figure 2 A sequence fusion convolutional network model structure in the embodiment is shown;

[0035] Figure 3 A sequence fusion convolutional network model construction process in the embodiment is shown;

[0036] Figure 4 A radiation source identification rate result in the embodiment is shown. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical scheme and advantages of the application more clear, specific and apparent, the application will be further described in detail below with reference to the drawings and technical scheme.

[0038] In order to solve the problems of single feature, poor feature fusion network design, ignoring sequence features and low recognition rate of the existing radiation source individual identification method, the embodiment of the application provides a specific radiation source identification method based on multi-sequence feature learning, fuses multiple sequence signal features and extracts spatial features and timing features in the multi-sequence features to improve the signal recognition efficiency, and the identification process can be specifically designed to include:

[0039] S101, generating I / Q road in-phase quadrature signal radiation source simulation signals of multiple modulation types according to a radiation source fingerprint feature generation mechanism, and using the radiation source simulation signals to build a sample data set.

[0040] Specifically, the simulation simulation signal can be designed to include the following contents:

[0041] First, according to the radiation source signal fingerprint feature generation mechanism, and combining the distortion model of the transmitter, the filter, the oscillator and the power amplifier distortion performance;

[0042] Then, according to the transmitter distortion model, the radiation source simulation signal is generated by adjusting the distortion model parameters.

[0043] According to the mechanism of the radiation source fingerprint characteristics, the distortion models of the quadrature modulator, filter, oscillator and power amplifier are given. The distortion models can be described as follows:

[0044] Due to the influence of various imperfect factors in the process of hardware production, the quadrature modulator will have the imbalance phenomenon of I / Q two-way signal waveform modulation, which is specifically manifested as gain mismatch, phase mismatch and DC bias. Assuming that s b,I (t) and s b,Q (t) are the baseband signal waveforms of I / Q two-way, the ideal baseband signal is shown in equation (1).

[0045] s0(t)=s b,I (t)+s b,Q (t) (1)

[0046] The baseband I / Q signal carrying the distortion of the I / Q modulator is shown in equation (2).

[0047] s(t)=(1-g)(s b,I (t)+c I )+j(1+g)(s b,Q (t)+c Q )e jφ (2)

[0048] Where g is the gain mismatch, φ is the phase deviation, c I and c Q are the DC components generated by the two-way mixer.

[0049] The distortion of the filter is mainly manifested as the tilt and ripple of the amplitude-frequency response, and the fluctuation of the group delay. Assuming that the ideal baseband shaping filter is g(t), the ideal transmission signal corresponding to it is shown in equation (3).

[0050]

[0051] In equation (3), f c is the carrier frequency, θ is the initial phase, τ is the time delay, {a k} is the symbol sequence, and T0 is the symbol period. Let G(f) be the frequency response of g t , then the frequency response of the distorted filter is shown in equation (4).

[0052] H(f)=G(f)A(f)e jφ(f) (4)

[0053] Where A(f)=a0+a n cos(2πα n f) is the amplitude distortion, and φ(f)=b0+b ncos(2πβ n f) is the phase distortion, a0and b0are linear gains, a n and b n are the fluctuation gains, a n and β n depend on the period of the amplitude ripples and the time delay fluctuations, respectively, then the signal carrying the filter distortion is given by equation (5).

[0054]

[0055] In equation (5), is the baseband shaping filter with distortion.

[0056] The oscillator distortion is mainly reflected in the phase noise around the carrier frequency. Based on the ideal signal in equation (3), let the phase noise be The signal carrying the oscillator distortion can be expressed as

[0057]

[0058] which is equivalent to adding a time-varying additive factor to the carrier frequency f c , which is usually characterized by a first-order autoregressive model as shown in equation (7).

[0059]

[0060] where v(t) is a Gaussian white noise with unit variance, c o reflects the individual differences of the transmitter. As can be seen from equation (7), when c o is larger, the randomness is stronger, and the disturbance to the carrier frequency f c is more obvious; on the contrary, the autocorrelation is larger, and the stability of the signal carrier frequency f c is higher.

[0061] The power amplifier distortion is mainly reflected in two aspects: the amplitude / phase compression effect, i.e. the signal amplitude is compressed in the saturation region; the amplitude / phase conversion effect, i.e. the signal amplitude in the non-saturation region produces additional phase. For narrowband signal power amplifier, Taylor series model is generally used to describe it. Assume that the ideal input of the power amplifier is as shown in equation (8), where p(t) is the ideal complex baseband waveform.

[0062]

[0063] Then, the signal carrying the power amplifier distortion is as shown in equation (9).​

[0064]

[0065] wherein {λ1, λ3, …, λ 2K-1} are coefficients of the Taylor series, λ1=1. Generally, λ3<0, and |λ k | decreases with the increase of k. Therefore, the second term of the formula (5)-(9) mainly represents the characteristics of λ3, thereby weakening the amplitude of the input signal, causing the AM / PM compression effect. When λ k is a complex number, the signal amplitude will be converted into additional phase, resulting in AM / PM conversion effect.

[0066] According to the distortion model, a plurality of radiation source signals are generated by adjusting the distortion model parameters. For example, 7 groups of different distortion model parameters are set to simulate 7 radiation sources. The radiation source simulation parameter settings are as follows: the source bit information is randomly generated, the modulation mode is QPSK, the symbol number is 200, the symbol rate is 500KBaud, the carrier frequency is 350KHz, the sampling rate is 10MHz, the filter is a raised cosine shaping filter with a roll-off factor of 0.35, the number of radiation sources is 7, and the radiation source signals are generated according to the above orthogonal transmitter distortion model, wherein the default {a0, b0, α n , β n}={1, 0, 4, 4}, and the remaining radiation source parameter distortion parameters are shown in Table 1.

[0067] Table 1 Distortion parameters of different radiation sources

[0068]

[0069] According to the simulation radiation source signal, signal sample data is assembled, each radiation source signal information point in the simulation signal is sampled, n information points collected continuously each time are combined to form a signal sample, and m signal samples of each modulation type radiation source simulation signal are collected; the signal sample data set is assembled according to all types of radiation source simulation signals, and the sample data set under each modulation type is divided according to a preset proportion, to obtain a training sample set, a verification sample set and a test sample set for network model training and optimization, wherein n and m are preset thresholds.

[0070] Specifically, the information points of each radiation source signal can be sampled at intervals of 8 information points, and 2000 information points are continuously collected each time to form a signal sample. For each radiation source signal, signals are collected at intervals of 2 dB between 0 dB and 30 dB, 500 samples are collected at each signal-to-noise ratio, and a total of 8000 samples are collected for each radiation source signal. All signals form a signal sample set, a total of 56000 samples. For each signal-to-noise ratio under each modulation pattern, 500 signal samples are collected, 60% of each modulation signal is extracted to form a training sample set, 20% of the remaining 40% is extracted to form a verification sample set, and the last 20% of the entire sample set is used as a test sample set.

[0071] In S102, the signals in the sample data set are standardized to obtain a multi-sequence signal composed of analog simulation signal I / Q sequences, amplitude and phase A / P sequences, and frequency domain A / P sequences.

[0072] Specifically, the signals in the sample data set are standardized, which can be designed to include the following contents:

[0073] First, each in-phase and quadrature signal in the signal sample is standardized to I / Q sequences, amplitude and phase A / P sequences, and frequency domain A / P sequences.

[0074] Next, the standardized signals are sampled to obtain a baseband signal complex sequence with a length of N, and the instantaneous amplitude and instantaneous phase of the signal are obtained according to the baseband signal complex sequence.

[0075] Then, the baseband signal complex sequence is subjected to Fourier transform to obtain a frequency domain complex sequence representation, and the amplitude spectrum and phase spectrum of the signal are obtained according to the frequency domain complex sequence representation.

[0076] Finally, the in-phase and quadrature signals in the signal sample and the corresponding instantaneous amplitude, instantaneous phase, amplitude spectrum, and phase spectrum are arranged to obtain a multi-sequence signal of the signal sample.

[0077] Each in-phase and quadrature (I / Q) signal sample is first standardized to I / Q sequences, amplitude and phase (A / P) sequences, and frequency domain A / P sequences, and the three sequences are combined together as multi-sequence features. The standardized received signal is sampled at a sampling rate F s After sampling, a baseband signal complex sequence with a length of N can be obtained, which can be specifically represented as

[0078] r(n) = r I (n) + j r Q (n), n = 0,..., N-1 (10)

[0079] I refers to the in-phase component, and Q refers to the quadrature component. The I path and the Q path are completely orthogonal. The instantaneous amplitude A(n) and the instantaneous phase P(n) of the signal can be defined as follows:

[0080]

[0081] The frequency domain complex sequence can be obtained by using Fourier transform, and is described as follows:

[0082]

[0083] The signal amplitude spectrum F(k) and the phase spectrum Φ(k) can be obtained by X(k), and are described as follows:

[0084]

[0085] The multi-sequence signal representation required in the scheme can be obtained by arranging the time domain sequence and the frequency domain A / P sequence, and is described as follows:

[0086]

[0087] S103, a sequence fusion convolutional network model for radiation source classification and identification is constructed, and the sequence fusion convolutional network model is trained and optimized by using the multi-sequence signals in the sample data set.

[0088] Specifically, the sequence fusion convolutional network model for radiation source classification and identification is constructed, and the model structure is as shown in Figure 2 The model structure can be designed to include: a feature fusion module for multi-sequence feature fusion, a squeeze-and-excitation module for recalibrating the sequence feature channel weight vector according to the sharing degree of the features in the input feature vector to generate a new feature vector, and a time sequence convolution classification module for classifying and identifying the new feature sequence.

[0089] The feature fusion module uses three multi-channel convolution blocks to perform parallel convolution and splicing on the I / Q sequence of the I / Q dual-channel sequence, the time domain A / P sequence, and the frequency domain A / P sequence, respectively, to form a multi-sequence signal feature.

[0090] The feature fusion module can be composed of three multi-channel convolution blocks and a splicing layer (Cancatenate). The multi-channel convolution block is composed of three parallel convolution units (Conv Unit1, Conv Unit2, and Conv Unit3), a splicing layer (Cancatenate1), and a convolution unit (Conv Unit4) in sequence. The convolution unit is composed of a one-dimensional convolution layer, a BN layer, and a ReLU activation function in sequence.

[0091] The structure of the multi-channel convolution block can be described as: one signal of the two-channel sequence → Conv Unit1 → Cancatenate1, another signal of the two-channel sequence → Conv Unit2 → Cancatenate1, the two-channel sequence → Conv Unit3 → Cancatenate1, and Cancatenate1 → Conv Unit4.

[0092] The structure of the feature fusion module can be described as: the time-domain I / Q sequence → multi-channel convolution block 1 → Cancatenate, the time-domain A / P sequence → multi-channel convolution block 2 → Cancatenate, the frequency-domain A / P sequence → multi-channel convolution block 3 → Cancatenate, and the combined multi-sequence feature → Cancatenate.

[0093] The squeezing excitation module is used for a feature vector X ∈ R L×C The global average pooling F sq (·) can generate a channel statistical vector Z ∈ R 1×C , where L is the length of the feature vector, C represents the number of channels of the feature vector, and the i-th element z i is calculated by the following formula, where L is the length of the feature vector, and C represents the number of channels of the feature vector.

[0094]

[0095] Z can be generated by a specific change F ex (·, W) of the sequence feature X channel weight vector S, where δ represents the ReLU function, and σ is the Sigmoid activation function, Each value in the weight vector represents the result of the SE block learning the importance of each channel of X to the classification task, and the greater the value, the more beneficial the channel is to the classification task.

[0096] S = F ex (Z, W) = σ (g (Z, W) ) = σ (W2δ (W1Z) ) (16)

[0097] After obtaining the sequence feature X channel weight vector S, S can be used to re-scale the channel weight of X by multiplication to generate a new feature vector The i-th element of the new feature vector is calculated by the following formula, where S = [s1, s2,..., s C ]·X. Compared with X, The weight of each channel is re-scaled according to its contribution to the classification task, so that the final classification task is more directional.

[0098]

[0099] The time series convolution classification module can be designed to be sequentially connected by eight time series residual blocks (TRB) and one fully connected output layer. The activation function of the fully connected output layer is Softmax.

[0100] The specific structure of the time series residual block can be described as: input → dilated convolution layer → normalization layer → ReLU function activation → random inactivation layer (Dropout) → dilated convolution layer → normalization layer → ReLU function activation → random inactivation layer (Dropout) → adder, input → 1x1 convolution layer → adder, adder → output.

[0101] The specific structure of the time series convolution classification module can be described as: input → TRB-1 → TRB-2 → TRB-3 →

[0102] TRB-4 → TRB-5 → TRB-6 → TRB-7 → TRB-8 → fully connected layer.

[0103] The specific parameters of each module in the network model are as follows:

[0104] In the three multi-channel convolution blocks, the number of convolution kernels of the single-channel convolution unit is 5, and the size of the convolution kernel is 7.

[0105] In the three multi-channel convolution blocks, the number of convolution kernels of the double-channel convolution unit is 10, and the size of the convolution kernel is 7.

[0106] In the three multi-channel convolution blocks, the number of convolution kernels of the convolution unit after the splicing layer is 20, and the size of the convolution kernel is 7.

[0107] The squeeze and excitation block is set to r=4.

[0108] The parameters of TRB-1 are: 64, 7, 1. It represents the dilated convolution layer of TRB-1, the number of convolution kernels is 64, the size of the convolution kernel is 7, and the dilation coefficient is 1.

[0109] The parameters of TRB-2 are: 64, 7, 2. It represents the dilated convolution layer of TRB-2, the number of convolution kernels is 64, the size of the convolution kernel is 7, and the dilation coefficient is 2.

[0110] The parameters of TRB-3 are: 64, 7, 4. It represents the dilated convolution layer of TRB-1, the number of convolution kernels is 64, the size of the convolution kernel is 7, and the dilation coefficient is 4.

[0111] The parameters of TRB-4 are: 64, 7, 8. It represents the dilated convolution layer of TRB-2, the number of convolution kernels is 64, the size of the convolution kernel is 7, and the dilation coefficient is 8.

[0112] The parameters of TRB-5 are: 64, 7, 16. It represents the dilated convolution layer of TRB-1, the number of convolution kernels is 64, the size of the convolution kernel is 7, and the dilated coefficient is 46.

[0113] The parameters of TRB-6 are: 64, 7, 32. It represents the dilated convolution layer of TRB-2, the number of convolution kernels is 64, the size of the convolution kernel is 7, and the dilated coefficient is 32.

[0114] The parameters of TRB-7 are: 64, 7, 64. It represents the dilated convolution layer of TRB-1, the number of convolution kernels is 64, the size of the convolution kernel is 7, and the dilated coefficient is 64.

[0115] The parameters of TRB-8 are: 64, 7, 128. It represents the dilated convolution layer of TRB-2, the number of convolution kernels is 64, the size of the convolution kernel is 7, and the dilated coefficient is 128.

[0116] The number of convolution kernels of the full connection layer is 7, corresponding to the number of output categories.

[0117] When training the sequence fusion convolution network with the training set, the Adam optimizer can be selected to optimize the network, the initial learning rate is set to 0.001, 64 samples are trained per batch, and the maximum training round of the entire training sample is 200. The model is verified using the verification set every training round, and the verification loss is used as a reference. When the verification loss does not decrease after 30 iterations, the training model is stopped.

[0118] Further, the arrangement order of all samples in the training sample can be shuffled during training, and the training sample and the verification sample are input into the sequence fusion convolution network model to perform iterative training of the network model. When the maximum number of network training rounds is reached or the early stopping mechanism condition is met, the training process of the neural network is completed, and the sequence fusion convolution network model after training optimization is obtained. The test sample set can be input into the trained multi-sequence fusion convolution network to obtain the recognition result, which is compared with the true category to statistically evaluate the recognition accuracy of the network model.

[0119] S104, input the to-be-identified radiation source signal into the sequence fusion convolution network model after training optimization, and obtain the modulation category of the to-be-identified radiation source signal by using the sequence fusion convolution network model after training optimization.

[0120] Further, based on the above method, the embodiment of the present application also provides a specific radiation source identification system based on multi-sequence feature learning, which comprises a data simulation module, a data processing module, a model construction module and a target identification module, wherein,

[0121] a data simulation module, configured to generate simulated signals of a signal source of a plurality of modulation types in I / Q paths according to a mechanism of generating a fingerprint feature of the signal source of the signal source, and to assemble a sample data set by using the simulated signals of the signal source;

[0122] a data processing module, configured to perform standardization processing on signals in the sample data set to obtain a multi-sequence signal composed of simulated signal I / Q sequences, amplitude / phase A / P sequences and frequency domain A / P sequences;

[0123] a model construction module, configured to construct a sequence fusion convolutional network model for classification and identification of the signal source and to train and optimize the sequence fusion convolutional network model by using the multi-sequence signal in the sample data set;

[0124] a target identification module, configured to input a signal of a signal source to be identified into the sequence fusion convolutional network model that is trained and optimized, and to obtain a modulation type of the signal of the signal source to be identified by using the sequence fusion convolutional network model that is trained and optimized.

[0125] To verify the effectiveness of the scheme, the following experimental data are further explained:

[0126] The simulation experiment is implemented on an NVIDIA Quadro RTX 6000 and a Keras2.6.0Tensorflow-GPU2.4.0 platform, and the generation of the modulation signal and the simulation experiment of the sequence fusion convolutional network model in the experimental scheme are completed. The sequence fusion convolutional network model adopts a lightweight dense convolutional long short-term memory network, and the parameters of the network are the specific parameters of each module of the network model in the embodiment. According to the steps (1) to (6) shown in the embodiment, the experiment is completed, and the specific signal source identification rate based on multi-sequence feature learning is obtained, as shown in FIG. 6. Figure 3 Figure 4 As shown in FIG. 6, the identification rate gradually increases and stabilizes with the increase of the signal-to-noise ratio, and the highest identification rate can reach 99.28%.

[0127] Through the above experimental data, it can be further verified that the scheme can improve the signal source identification rate and efficiency by fully fusing a plurality of sequence features and extracting spatial features and time sequence features in the multi-sequence features, and is convenient for the application in actual scenes such as signal detection and / or defense and electronic countermeasures.

[0128] Unless otherwise specifically stated, the relative steps, numerical expressions and values of the components and steps set forth in these embodiments do not limit the scope of the present application.

[0129] ​The various embodiments are described in the specification in a progressive manner, each embodiment focusing on the differences from other embodiments, and the same or similar parts between embodiments can be mutually referred to. For the system disclosed by the embodiments, since it corresponds to the method disclosed by the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0130] The units and method steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been described in the above description in general terms. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation does not exceed the scope of the present application.

[0131] Those skilled in the art can understand that all or part of the steps of the above method can be instructed by a program to complete the relevant hardware, and the program can be stored in a computer readable storage medium, such as a read-only memory, a magnetic disk or an optical disk, etc. Alternatively, all or part of the steps of the above embodiments can also be implemented using one or more integrated circuits, and accordingly, each module / unit in the above embodiments can be implemented in the form of hardware or in the form of a software function module. The present application is not limited to any specific form of combination of hardware and software.

[0132] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present application, which are used to illustrate the technical solutions of the present application, and are not limiting. The protection scope of the present application is not limited thereto, although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily think of changes to the technical solutions recorded in the foregoing embodiments within the technical scope disclosed by the present application, or make equivalent replacements to some of the technical features; and these modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for identifying specific radiation sources based on multi-sequence feature learning, characterized in that, Include: Based on the generation mechanism of radiation source fingerprint features, I / Q channel in-phase orthogonal signal radiation source simulation signals of various modulation types are generated, and sample datasets are constructed using radiation source simulation signals. The signals in the sample dataset are standardized to obtain a multi-sequence signal composed of simulated signal I / Q sequence, amplitude-phase A / P sequence, and frequency domain A / P sequence. A sequence fusion convolutional network model for radiation source classification and identification was constructed, and the model was trained and optimized using multiple sequence signals from the sample dataset. The radiation source signal to be identified is input into the trained and optimized sequence fusion convolutional network model, and the modulation category of the radiation source signal to be identified is obtained by using the trained and optimized sequence fusion convolutional network model. The constructed sequence fusion convolutional network model for radiation source classification and identification includes: a feature fusion module for multi-sequence feature fusion, a squeezing excitation module for recalibrating the sequence feature channel weight vectors based on the degree of feature sharing in the input feature vector to generate new feature vectors, and a temporal convolutional classification module for classifying and identifying new feature sequences. The feature fusion module uses three multi-channel convolutional blocks to perform parallel convolution and splicing on the I / Q sequence, time-domain A / P sequence, and frequency-domain A / P sequence of the I / Q dual-channel sequence, respectively, to form multi-sequence signal features; The process by which the squeezing excitation module recalibrates the sequence feature channel weight vector based on the degree of feature sharing in the classification task of the input feature vector to generate a new feature vector is represented as follows: in, For the new feature vector The i-th element, s i The i-th element X in the channel weight vector S of the input feature vector X. i The channel weights are S = σ(W2δ(W1Z)), where Z is the channel statistics vector generated by global average pooling of the output feature vector of the first layer on the network, δ is the ReLU function, σ is the Sigmoid activation function, W1 is the weight parameter of the first fully connected layer in the squeeze activation module, and W2 is the weight parameter of the second fully connected layer in the squeeze activation module.

2. The method for identifying specific radiation sources based on multi-sequence feature learning according to claim 1, characterized in that, Based on the mechanism of radiation source fingerprint feature generation, various modulation types of radiation source simulation signals are generated, including: First, based on the generation mechanism of the fingerprint characteristics of radiation source signals, and combined with the distortion performance of quadrature modulators, filters, oscillators and power amplifiers, a transmitter distortion model is constructed. Then, based on the transmitter distortion model, the radiation source simulation signal is generated by adjusting the distortion model parameters.

3. The method for identifying specific radiation sources based on multi-sequence feature learning according to claim 1 or 2, characterized in that, A sample dataset is constructed using simulated radiation source signals. This includes: sampling signal information points for each type of radiation source in the simulated signals, combining n consecutively collected information points into a signal sample, and collecting m signal samples for each type of modulation of the simulated radiation source signals; constructing a sample dataset based on the signal samples of all types of simulated radiation source signals, and dividing the sample dataset for each modulation type according to a preset ratio to obtain a training sample set, a validation sample set, and a test sample set for network model training and optimization, where n and m are preset thresholds.

4. The method for identifying specific radiation sources based on multi-sequence feature learning according to claim 1, characterized in that, Standardization of signals in the sample dataset includes: First, each in-phase orthogonal signal in the signal sample is normalized into an I / Q sequence, an amplitude-phase (A / P) sequence, and a frequency domain (A / P) sequence; Next, the standardized signal is sampled to obtain a complex sequence of baseband signal of length N, and the instantaneous amplitude and instantaneous phase of the signal are obtained based on the baseband signal complex sequence. Then, a Fourier transform is performed on the complex sequence of the baseband signal to obtain a complex sequence representation in the frequency domain, and the amplitude spectrum and phase spectrum of the signal are obtained based on the complex sequence representation in the frequency domain. Finally, the in-phase orthogonal signals in the signal sample and their corresponding instantaneous amplitude, instantaneous phase, amplitude spectrum and phase spectrum are arranged to obtain the multi-sequence signal of the signal sample.

5. The method for identifying specific radiation sources based on multi-sequence feature learning according to claim 4, characterized in that, Multi-sequence signals are represented as: Where, r I (), r Q (), A(), P(), F(), These represent the in-phase component, quadrature component, instantaneous amplitude, instantaneous phase, amplitude spectrum, and phase spectrum in the complex sequence of the baseband signal, respectively, and N is the length of the complex sequence of the baseband signal.

6. The method for identifying specific radiation sources based on multi-sequence feature learning according to claim 1, characterized in that, The temporal convolutional classification module classifies and identifies new feature sequences, including: First, information in the new feature sequence is extracted using several temporal residual blocks that are connected end to end to the input and output. Each temporal residual block extracts the temporal information of the input features and obtains the output features based on the input features and the extracted temporal information. Then, the output features of the final temporal residual block are classified using the Softmax activation function of the fully connected layer to obtain the radiation source signal type.

7. A specific radiation source identification system based on multi-sequence feature learning, characterized in that, It includes: a data simulation module, a data processing module, a model building module, and a target recognition module. The data simulation module is used to generate I / Q channel in-phase orthogonal signal radiation source simulation signals of various modulation types based on the generation mechanism of radiation source fingerprint characteristics, and to construct sample datasets using radiation source simulation signals; The data processing module is used to standardize the signals in the sample data to obtain a multi-sequence signal composed of the simulated signal I / Q sequence, amplitude-phase (A / P) sequence, and frequency domain (A / P) sequence. The model building module is used to build a sequence fusion convolutional network model for radiation source classification and identification, and to train and optimize the sequence fusion convolutional network model using multiple sequence signals in the sample dataset. The target recognition module is used to input the radiation source signal to be identified into the trained and optimized sequence fusion convolutional network model, and use the trained and optimized sequence fusion convolutional network model to obtain the modulation category of the radiation source signal to be identified. The constructed sequence fusion convolutional network model for radiation source classification and identification includes: a feature fusion module for multi-sequence feature fusion, a squeezing excitation module for recalibrating the sequence feature channel weight vectors based on the degree of feature sharing in the input feature vector to generate new feature vectors, and a temporal convolutional classification module for classifying and identifying new feature sequences. The feature fusion module uses three multi-channel convolutional blocks to perform parallel convolution and splicing on the I / Q sequence, time-domain A / P sequence, and frequency-domain A / P sequence of the I / Q dual-channel sequence, respectively, to form multi-sequence signal features; The process by which the squeezing excitation module recalibrates the sequence feature channel weight vector based on the degree of feature sharing in the classification task of the input feature vector to generate a new feature vector is represented as follows: in, For the new feature vector The i-th element, s i The i-th element X in the channel weight vector S of the input feature vector X. i The channel weights are S = σ(W2δ(W1Z)), where Z is the channel statistics vector generated by global average pooling of the output feature vector of the first layer on the network, δ is the ReLU function, σ is the Sigmoid activation function, W1 is the weight parameter of the first fully connected layer in the squeeze activation module, and W2 is the weight parameter of the second fully connected layer in the squeeze activation module.

Images