A communication interference identification method based on three-dimensional time-frequency and deep migration learning

Abstract

The present application relates to a kind of communication interference identification methods based on three-dimensional time-frequency and depth migration learning, comprising: S1, data preprocessing;S2, migration learning feature extraction;S3, LSTM time sequence feature extraction;S4, time-frequency chart feature, time sequence feature weight distribution realizes interference signal identification.The present application processes original spectrum data into three-dimensional time-frequency chart, retains multidimensional data information;Through the way of migration learning, the feature of three-dimensional time-frequency chart is extracted, the problem of small number of interference data samples is solved;Combined with LSTM network, the time sequence correlation feature of feature map is extracted;Finally, in combination with self-attention mechanism network (Self-Attention Mechanism), three-dimensional time-frequency chart feature, time sequence feature is assigned weight, it is helpful to model identification key feature and improve identification precision.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wireless communication, and particularly relates to a communication interference recognition method based on three-dimensional time-frequency and deep transfer learning. BACKGROUND

[0002] With the development of wireless communication technology, communication systems are facing an increasingly complex interference environment. The recognition and classification of interference signals are crucial for ensuring communication quality.

[0003] Currently, the traditional time-frequency analysis method commonly used for communication interference signal recognition, such as short-time Fourier transform and wavelet transform, can provide time-frequency representation of signals, but is limited in processing non-stationary signals or capturing three-dimensional characteristics of signals.

[0004] Therefore, in order to better capture the time-space characteristics of signals, the application aims to develop a communication interference recognition method based on three-dimensional time-frequency and deep transfer learning. SUMMARY

[0005] The application aims to overcome the shortcomings of the prior art and provide a communication interference recognition method based on three-dimensional time-frequency and deep transfer learning. The original spectrum data is processed into a three-dimensional time-frequency graph, retaining multi-dimensional data information. The features of the three-dimensional time-frequency graph are extracted through transfer learning, solving the problem of small number of interference data samples. The LSTM network is combined to extract the time sequence correlation features of the feature map. Finally, the Self-Attention Mechanism network is combined to assign weights to the three-dimensional time-frequency graph features and time sequence features, which helps the model to identify key features and improve the communication interference recognition accuracy.

[0006] The application solves its technical problem through the following technical scheme:

[0007] A communication interference recognition method based on three-dimensional time-frequency and deep transfer learning, the steps of the method are:

[0008] S1, data preprocessing: after time-frequency conversion, frequency point energy calculation and amplitude adjustment of the original IQ signal, a three-dimensional time-frequency graph is obtained through data fusion;

[0009] S2, feature extraction by transfer learning: input the three-dimensional time-frequency graph into the pre-trained Resnet neural network for image feature extraction;

[0010] S3, LSTM time sequence feature extraction: the obtained image features are converted in dimension and input into the LSTM network to obtain time sequence features;

[0011] S4, input the image features obtained in S2 and the time sequence features obtained in S3 into a Transformer network, perform classification through Linear and log_softmax, and complete the interference signal recognition.

[0012] Moreover, S1 is specifically:

[0013] 1) Fast Fourier transform processing: the original IQ signal is processed through fast Fourier transform to complete time-frequency conversion. The fast Fourier transform formula of the original IQ signal for a vector s(n) containing N uniformly sampled points is:

[0014]

[0015] Wherein:

[0016] k=0~2 N -1;

[0017] After fast Fourier transform, F n is the signal frequency represented by any sampling point n:

[0018]

[0019] Wherein: Fs is the sampling frequency;

[0020] 2) Frequency point energy calculation: the input and output of fast Fourier transform are both IQ complex signal data, and there is a 90-degree phase difference, representing the real part and imaginary part of the complex signal respectively. The frequency point energy calculation takes the modulus of the real part and imaginary part of each frequency point signal to obtain the projection of the signal vector in the real number domain, i.e. the signal energy,

[0021] The data output after fast Fourier transform is Z(n)=a+bi(a,b∈R), and the specific calculation formula of frequency point energy is:

[0022]

[0023] 3) Amplitude adjustment: taking the logarithm of the frequency point energy facilitates effective observation of signal changes, and the calculation formula is:

[0024] P n =20log|z n |,

[0025] 4) Forming frequency-power data Pf n by the frequency F n and energy P n corresponding to the sampling point n:

[0026] 5) Generating three-dimensional time-frequency graph: generating a three-dimensional time-frequency graph by continuously processing the Pfn The data is truncated according to a fixed time window W, and the data of the continuous N time windows is processed in the time dimension to generate three-dimensional time-frequency graph data x, and the processing formula is:

[0027] x = [Pf n ] W,N = [Pf n ] W ×N

[0028] Wherein: [Pf n ] W The continuous frequency-power data Pf n is truncated according to the time window W to obtain the frequency-power data in each time window;

[0029] ×N indicates that the data of N time windows is stacked in the time dimension to form a three-dimensional time-frequency graph;

[0030] The dimension of x is (frequency f, power P, time t).

[0031] Moreover, S2 is specifically: the feature extraction selects a Resnet neural network pre-trained on an ImageNet dataset, and x(n) after data preprocessing is a [b, c, h, w] dimensional vector, which is input into the pre-trained Resnet neural network, wherein b is batch size, c is channel number, h is height, and w is width. The feature obtained after passing through the Resnet neural network is Resnet out = Resnet(x); wherein Resnet out ∈R b×c′×h′×w′ ,

[0032] Resnet(x) = F(x, {W i}) + x

[0033] Wherein F(x, {W i}) is a convolution operation in a residual block, and W i is a weight parameter of a convolution layer. +x indicates a skip connection in residual calculation.

[0034] Moreover, S3 is specifically:

[0035] 1) Resnet out is converted in dimension according to the following formula:

[0036]

[0037] Wherein: T = h' x w' indicates that the h' x w' two dimensions of the data Resnet out ∈R b×c′×h′×w′ are spliced;

[0038] 2) The LSTM (input) input into the LSTM(x) network to obtain the time sequence feature, the calculation of LSTM(x) is as follows

[0039] The formula is as follows:

[0040] i t = σ (w ii x t +b ii +w hi h t-1 +b hi ) ;

[0041] f t = σ (w if x t +b if +w hf h t-1 +b hf ) ;

[0042] g t = tanh (w ig x t +b ig +w hg h t-1 +b hg ) ;

[0043] o t = σ (w io x t +b io +w ho h t-1 +b ho ) ;

[0044] c t = f t ⊙c t-1 +i t ⊙g t ;

[0045] h t = o t ⊙tanh (c t ) ;

[0046] Wherein: i t , f t , g t , o t are input gate, forget gate, candidate cell state and output gate respectively;

[0047] c t is the cell state; h tis the hidden state, and is the sigmoid activation function; represents element-wise multiplication.

[0048] After the above formula calculation output: where d lstm is the dimension of the LSTM network hidden layer.

[0049] Moreover, the S4 is specifically:

[0050] 1) Resnet out and LSTM (output) as input data SAM (input) into the self-attention mechanism network, wherein the self-attention mechanism network is composed of a multi-head attention module and an encoder module,

[0051] The self-attention mechanism network calculation process is as follows:

[0052] Q = SAM (input) W Q , K = SAM (input) W K , V = SAM (input) W V ;

[0053]

[0054] SAM (output) = Embedding (Attention (Q, A, W));

[0055] Wherein: Q, K, V are query (Query), key (Key) and value (Value) matrix respectively;

[0056] W Q , W K , W V are learnable weight matrices;

[0057] Embedding encoder module: composed of Linear layer, Dropout layer, LaryerNorm layer;

[0058] 2) Interference type identification: composed of Linear and log_softmax, the probability of each category is calculated by the function log_softmax, and the interference signal identification is completed;

[0059] The calculation formula of the interference signal identification network is:

[0060] Y = log_softmax (linear (SAM out , C))

[0061] Where: Y ∈ Rb×c is a classification result;

[0062] C is the number of categories.

[0063] The positive effects that can be produced by the present application are:

[0064] 1. The present application provides more abundant input data for a deep learning model by constructing a three-dimensional time-frequency graph including time, frequency, energy, and spatial distribution information; and combines transfer learning to migrate pre-trained network parameters to a communication interference identification task, thereby improving the performance of the model on a small sample data set.

[0065] 2. The present application processes original spectrum data into a three-dimensional time-frequency graph, retains multi-dimensional data information, extracts features of the three-dimensional time-frequency graph through transfer learning, solves the problem of a small number of interference data samples, combines an LSTM network to extract time sequence correlation features of a feature map, and finally combines an attention mechanism of a Transformer network to assign weights to three-dimensional time-frequency graph features and time sequence features, which is helpful for the model to identify key features and improve communication interference identification accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 is a flowchart of data preprocessing of the present application;

[0067] Figure 2 is a two-dimensional time-frequency diagram of different interferences at a fixed time point after FFT, frequency point energy calculation, and amplitude adjustment of the present application;

[0068] Figure 3 is a three-dimensional time-frequency diagram generated after data fusion of the present application, which has time, frequency, energy, and spatial distribution information;

[0069] Figure 4 is a network structure diagram of the present application;

[0070] Figure 5 is a network training convergence curve diagram of the present application;

[0071] Figure 6 is a recognition accuracy confusion matrix diagram of the network of the present application. DETAILED DESCRIPTION

[0072] The present application will be further described in detail through specific embodiments below, which are only descriptive and not limiting, and cannot limit the protection scope of the present application.

[0073] A communication interference identification method based on three-dimensional time-frequency and deep transfer learning, which is characterized in that the steps of the method are:

[0074] S1, data preprocessing: after the original IQ signal completes time-frequency conversion, frequency point energy calculation and amplitude adjustment, a three-dimensional time-frequency graph is obtained through data fusion;

[0075] 1) Fast Fourier transform processing: the original IQ signal is processed by fast Fourier transform to complete time-frequency conversion. The fast Fourier transform formula of the original IQ signal for a vector s(n) containing N uniformly sampled points is:

[0076]

[0077] Wherein:

[0078] k = 0 ~ 2 N -1;

[0079] After fast Fourier transform, F n is the signal frequency represented by any sampling point n:

[0080]

[0081] Wherein: Fs is the sampling frequency;

[0082] 2) Frequency point energy calculation: the input and output of fast Fourier transform are both IQ complex signal data, and there is a 90-degree phase difference, which represents the real part and imaginary part of the complex signal respectively. The frequency point energy calculation takes the modulus of the real part and imaginary part of each frequency point signal to obtain the projection of the signal vector in the real number domain, that is, the signal energy,

[0083] The data output after fast Fourier transform is Z(n) = a + bi (a, b ∈ R), and the specific calculation formula of frequency point energy is:

[0084]

[0085] 3) Amplitude adjustment: taking the logarithm of the frequency point energy facilitates effective observation of signal changes, and the calculation formula is:

[0086] P n = 20log|z n |,

[0087] 4) Forming frequency point-power data Pf n by the frequency F n and the energy P n corresponding to the sampling point n:

[0088] 5) Generating a three-dimensional time-frequency graph: generating a three-dimensional time-frequency graph by continuously processing Pf nThe data is truncated according to a fixed time window W, and then the data from N consecutive time windows are processed in three dimensions according to the time dimension to generate a three-dimensional time-frequency plot data x. The processing formula is:

[0089] x = [Pf] n ] W,N =[Pf n W×N

[0090] Among them: [Pf n ] W This indicates that the continuous frequency-power data Pf n The frequency-power data within each time window is obtained by truncating the data according to the time window W.

[0091] ×N represents stacking the data from N time windows along the time dimension to form a three-dimensional time-frequency graph;

[0092] The dimensions of x are (frequency f, power P, time t).

[0093] like Figure 1 The original IQ signal data is processed through time-frequency conversion, frequency point energy calculation, and amplitude adjustment to obtain the following: Figure 2 The two-dimensional time-frequency diagrams of different interferences shown are from... Figure 2 The two-dimensional time-frequency plot provides a relatively objective display of the spectral characteristics of the interference signal that do not change over time at the current time point, such as... Figure 2 b) random noise interference, c) broadband interference, and e) comb interference are clearly distinguishable on a two-dimensional time-frequency diagram. However, a) swept frequency interference and d) fixed frequency interference cannot be distinguished by a time-frequency diagram at a fixed time.

[0094] Will Figure 2 Two-dimensional time-frequency data are generated through data fusion as follows Figure 3 The three-dimensional time-frequency diagram shown contains information on time, frequency, energy, and spatial distribution, such as... Figure 3 The frequency sweeping interference characteristic shown is a diagonal line that is discrete in time and varies in frequency intervals, while fixed frequency interference is a continuous interference signal at a fixed frequency point. The characteristics of the two are obvious.

[0095] S2. Feature extraction through transfer learning: Input the 3D time-frequency graph into the pre-trained ResNet neural network to extract image features;

[0096] The feature extraction uses a ResNet neural network pre-trained on the ImageNet dataset. After data preprocessing, x(n) is a [b, c, h, w]-dimensional vector, which is input into the pre-trained ResNet neural network, where b is the batch size, c is the number of channels, h is the height, and w is the width. The features obtained after passing through the ResNet neural network are...out = Resnet(x); wherein Resnet out ∈R b×c′×h′×w′ .

[0097] S3, LSTM time sequence feature extraction: the obtained image features are dimensionally converted and input into an LSTM network to obtain time sequence features;

[0098] 1) Resnet out is dimensionally converted according to the following formula:

[0099]

[0100] wherein: T = h' * w';

[0101] 2) LSTM (input) is input into an LSTM network to obtain time sequence features:

[0102] LSTM (output) = LSTM(LSTM (input) );

[0103] wherein d lstm is the dimension of the hidden layer of the LSTM network.

[0104] S4, the image features obtained in S2 and the time sequence features obtained in S3 are input into a self-attention mechanism network (Self-Attention Mechanism) as input data, classification is performed through Linear and log_softmax, and interference signal recognition is completed;

[0105] 1) The self-attention mechanism network (Self-Attention Mechanism) is composed of a multi-head attention module (MultiheadAttention) and an encoder module (Embedding). The Embedding encoder module is composed of a Linear layer, a Dropout layer and a LaryerNorm layer;

[0106] 2) Interference type recognition: composed of Linear and log_softmax, the probability of each category is calculated through the function log_softmax, and interference signal recognition is completed.

[0107] As Figure 4 is a network structure schematic diagram used by the present application, wherein:

[0108] Feature extraction of transfer learning: input the three-dimensional time-frequency diagram generated in S1 into the pre-trained Resnet neural network for image feature extraction; in the actual implementation process, the pre-trained network can select Resnet152, Resnet50 or AlexNet, Googlenet pre-trained network, and the present application is not limited to the residual network;

[0109] LSTM time sequence feature extraction: the feature data output by the Resnet neural network needs to be converted in dimension through a View module, and then the time sequence feature is extracted through an LSTM module.

[0110] The image features obtained in S2 and the time sequence features obtained in S3 are input into a self-attention mechanism network (Self-Attention Mechanism) as input data, classified through Linear and log_softmax, and the interference signal recognition is completed;

[0111] The self-attention mechanism network (Self-Attention Mechanism) is composed of a multi-head attention module (MultiheadAttention) and an encoder module (Embedding). The Embedding encoder module is composed of a Linear layer, a Dropout layer and a LaryerNorm layer;

[0112] Interference type recognition: composed of Linear and log_softmax, the probability of each category is calculated through the function log_softmax, and the interference signal recognition is completed.

[0113] As Figure 5 , 6 The training convergence curve and interference recognition effect diagram of the present application are shown in the figures, and it can be seen from the figures that:

[0114] (1) After the data preprocessing process of the present application, a total of 1000 three-dimensional time-frequency data of different interference types are generated, and in the implementation process of the present application, 400 data are used as a training set, 300 data are used as a validation set, and 300 data are used as a test set. The convergence curve under the condition that the pre-trained residual network is Resnet152 is shown in Figure 5 .

[0115] (2) As shown in Figure 5 , after 50 epochs of training, the loss of the training data set and the validation data set is reduced to below 0.1, and the model is successfully converged.

[0116] (3) Figure 6Confusion matrix of recognition results of different interferences on the test data set. Among them, labels 0-5 respectively represent six types of interference, i.e. no interference, sweep interference, random noise interference, wideband interference, comb interference and fixed frequency interference. The recognition accuracy for no interference is 92%, which is relatively low. The main reason is that there are other device product spectrum interferences in the space. The recognition accuracy under the condition of interference is above 98%.

[0117] (4) The present application adopts the way of transfer learning, which is very suitable for the case of a small number of interference samples in actual application scenarios. Compared with not using transfer learning, using 400 training data sets, the feature extraction Resnet152 network in the network structure of the present application cannot obtain a better recognition accuracy.

[0118] Although the embodiments of the present application and the drawings are disclosed for the purpose of illustration, those skilled in the art can understand that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present application and the appended claims, therefore, the scope of the present application is not limited to the content disclosed in the embodiments and drawings.

Claims

1. A method for communication jamming identification based on three-dimensional time-frequency and depth migration learning, characterized in that: The steps of the method are: S1, data preprocessing: after completing time-frequency conversion, frequency point energy calculation and amplitude adjustment of the original IQ signal, a three-dimensional time-frequency graph is obtained through data fusion; S2, feature extraction of transfer learning: input the three-dimensional time-frequency graph into the pre-trained Resnet neural network for image feature extraction; S3, LSTM time sequence feature extraction: the obtained image features are converted in dimension and input into the LSTM network to obtain time sequence features; S4, input the image features obtained in S2 and the time sequence features obtained in S3 into the self-attention mechanism network, and then combine Linear and log_softmax after weight allocation through the attention module to complete the interference signal recognition; The S1 is specifically: 1) Fast Fourier Transform processing: the original IQ signal is processed by fast Fourier transform to complete time-frequency conversion, and a vector s containing N uniformly sampled points is obtained The fast Fourier transform formula of the original IQ signal is: Z ； wherein: ; ； After a fast Fourier transform, The signal frequency represented by any one sampling point n: ； wherein: is the sampling frequency; 2) frequency point energy calculation: the input and output of fast Fourier transform are both IQ complex signal data, and there is a 90-degree phase difference, which represents the real part and imaginary part of the complex signal respectively. The frequency point energy calculation calculates the real part and imaginary part of each frequency point signal, and obtains the projection of the signal vector in the real number domain, that is, the signal energy, The data output after the fast Fourier transform is The frequency point energy is specifically calculated according to the following formula: ； 3) amplitude adjustment: take the logarithm of the frequency point energy, and the calculation formula is: ； 4) the frequency corresponding to the sampling point n Forming frequency-power data: , 5) Generate a three-dimensional time-frequency graph: This involves processing the continuous... The data is truncated according to a fixed time window W, and then the data from N consecutive time windows are processed in three dimensions according to the time dimension to generate a three-dimensional time-frequency plot. The calculation formula is: ×N； wherein: represents truncating the continuous frequency-power data truncating according to a time window W to obtain frequency-power data in each time window; x represents the data of N time windows stacked in the time dimension to form a three-dimensional time-frequency graph; The dimension of x is: frequency f, power P, time t.

2. The method of claim 1, wherein the method is based on a three-dimensional time-frequency and depth migration learning-based communication interference identification method. The S2 is specifically: the feature extraction selects a Resnet neural network pre-trained on an ImageNet dataset, and the data preprocessed is a [b, c, h, w] dimensional vector, b is the batch size, c is the channel number, h is the height, and w is the width, and the features obtained after the Resnet neural network are = ; wherein ∈ , +x； wherein: is a convolution operation in the residual block; is a weight parameter of the convolution layer; +x denotes a skip connection in the residual calculation.

3. The method of claim 1, wherein the method further comprises: The S3 is specifically: 1) To The dimension conversion is done according to the following formula: = (b, , T); Where: T= Indicates data ∈ of The two dimensions are spliced ​​together. 2) the input is fed into the LSTM(x) network to obtain the time-series features, and the calculation formula of the LSTM(x) is as follows: Input into the LSTM(x) network to obtain time-series features, and the calculation formula of the LSTM(x) is as follows: + ； + ； + ； + ； ⊙ ； ⊙ ( )； wherein: , , , are the input gate, the forget gate, the candidate cell state, and the output gate, respectively; is the unit state; is a hidden state; is a sigmoid activation function; ⊙ represents element-wise multiplication; After the above formula is calculated, output: ∈ wherein is the dimension of the LSTM network hidden layer.

4. The method of claim 1, wherein the method further comprises: The S4 is specifically: 1) To As input data into a self-attention mechanism network, which consists of a multi-head attention module and an encoder module, the self-attention mechanism network calculation process is as follows: ， ， ； MultiheadAttention ; ； Wherein: Q, K, V are query (Query), key (Key) and value (Value) matrices respectively; , , is a learnable weight matrix; The Embedding encoder module is composed of a Linear layer, a Dropout layer and a LaryerNorm layer; 2) interference type recognition: composed of Linear and log_softmax, the probability of each category is calculated through the function log_softmax to complete the interference signal recognition; The calculation formula of the interference signal recognition network is: ； where: Y e is the classification result; C is the number of categories.

Images