Communication interference identification method based on three-dimensional time frequency and deep transfer learning

Abstract

The invention relates to a communication interference identification method based on three-dimensional time frequency and deep transfer learning. The method comprises the following steps: S1, data preprocessing; s2, feature extraction of transfer learning; s3, LSTM time sequence feature extraction is carried out; and S4, carrying out time-frequency graph feature and time sequence feature weight distribution to realize interference signal identification. According to the method, original frequency spectrum data is processed into a three-dimensional time-frequency diagram, and multi-dimensional data information is reserved; features of the three-dimensional time-frequency graph are extracted in a transfer learning mode, and the problem that the number of interference data samples is small is solved; in combination with an LSTM network, extracting time sequence relevance features of the feature map; and finally, in combination with a self-attention mechanism network (Self-Attention Mode), weights are distributed for the three-dimensional time-frequency graph features and the time sequence features, which is beneficial for the model to identify key features and improve the identification precision.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications, and in particular relates to a communication interference identification method based on three-dimensional time-frequency and deep transfer learning. Background Art

[0002] With the development of wireless communication technology, communication systems are facing increasingly complex interference environments. The identification and classification of interference signals are crucial to ensuring communication quality.

[0003] At present, traditional time-frequency analysis methods such as short-time Fourier transform and wavelet transform are commonly used to identify communication interference signals. These two transforms can provide time-frequency representation of signals, but are limited in processing non-stationary signals or capturing the three-dimensional characteristics of signals.

[0004] Therefore, in order to better capture the spatiotemporal characteristics of the signal, the present invention aims to develop a communication interference identification method based on three-dimensional time-frequency and deep transfer learning. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a communication interference identification method based on three-dimensional time-frequency and deep transfer learning, which processes the original spectrum data into a three-dimensional time-frequency graph, retains multi-dimensional data information, and extracts the features of the three-dimensional time-frequency graph by transfer learning to solve the problem of a small number of interference data samples; combines with an LSTM network to extract the temporal correlation features of the feature graph; and finally combines with a self-attention mechanism network to assign weights to the three-dimensional time-frequency graph features and temporal features, which helps the model identify key features and improves the accuracy of communication interference identification.

[0006] The present invention solves the technical problem by the following technical solutions:

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

[0008] S1. Data preprocessing: After completing the time-frequency conversion, frequency energy calculation and amplitude adjustment of the original IQ signal, a three-dimensional time-frequency diagram is obtained through data fusion;

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

[0010] S3, LSTM time series feature extraction: The obtained image features are transformed into dimensions and input into the LSTM network to obtain time series features;

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

[0012] Moreover, the S1 is specifically:

[0013] 1) Fast Fourier transform processing: The original IQ signal is processed by fast Fourier transform to complete the time-frequency conversion. For a vector s(n) containing N uniform sampling points, the fast Fourier transform formula of the original IQ signal is:

[0014]

[0015] in:

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

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

[0018]

[0019] Where: Fs is the sampling frequency;

[0020] 2) Frequency point energy calculation: The input and output of the fast Fourier transform are both IQ complex signal data with a phase difference of 90 degrees, representing the real and imaginary parts of the complex signal respectively. The frequency point energy calculation performs modulo calculation on the real and imaginary parts of each frequency point signal to obtain the projection of the signal vector in the real number domain, that is, the signal energy.

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

[0022]

[0023] 3) Amplitude adjustment: Take the logarithm of the frequency energy to effectively observe signal changes. The calculation formula is:

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

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

[0026] 5) Generate three-dimensional time-frequency diagram: The continuous Pfn The data is truncated according to a fixed time window W, and then the data of N consecutive time windows are processed in three dimensions according to the time dimension to generate a three-dimensional time-frequency graph data x. The processing formula is:

[0027] x=[Pf n ] W,N =[Pf n ] W ×N

[0028] Where: [Pf n ] W Indicates that the continuous frequency-power data Pf n Truncate according to the time window W to obtain the frequency-power data in each time window;

[0029] ×N means stacking the data of N time windows in the time dimension to form a three-dimensional time-frequency diagram;

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

[0031] Moreover, S2 is specifically as follows: the feature extraction selects the Resnet neural network pre-trained on the ImageNet dataset, and the x(n) after data preprocessing 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 feature obtained after the Resnet neural network is Resnet out =Resnet(x); where Resnet out ∈R b×c′×h′×w′ ,

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

[0033] Where F(x,{W i}) is the convolution operation in the residual block, W i is the weight parameter of the convolutional layer. +x represents the skip connection in the residual calculation.

[0034] Moreover, the S3 is specifically:

[0035] 1) Resnet out The dimension conversion is performed according to the following formula:

[0036]

[0037] Where: T = h'×w' means that the data Resnet out ∈R b×c′×h′×w′ The h'×w' dimensions are spliced ​​together;

[0038] 2) LSTM (input) Input into the LSTM(x) network to obtain the time series features. The calculation formula of LSTM(x) is

[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] Where: i t 、f t , g t , o t They are input gate, forget gate, candidate unit state and output gate;

[0047] c t is the unit state; h tis the hidden state, σ is the sigmoid activation function; ⊙ represents element-wise multiplication;

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

[0049] Moreover, the S4 is specifically:

[0050] 1) Resnet out and LSTM (output) As input data SAM (input) to the self-attention mechanism network, where the self-attention mechanism network consists of a multi-head attention module and an encoder module.

[0051] The calculation process of the self-attention mechanism network 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] Where: Q, K, V are query, key, and value matrices respectively;

[0056] W Q , W K , W V is a learnable weight matrix;

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

[0058] 2) Interference type identification: It consists of Linear and log_softmax. The probability of each category is calculated through the log_softmax function to complete the interference signal identification;

[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 the classification result;

[0062] C is the number of categories.

[0063] The positive effects that the present invention can produce are:

[0064] 1. The present invention provides richer input data for deep learning models by constructing a three-dimensional time-frequency diagram, including time, frequency, energy, and spatial distribution information; at the same time, combined with transfer learning, the pre-trained network parameters are transferred to the communication interference identification task to improve the performance of the model on a small sample data set.

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

[0066] Figure 1 It is a schematic diagram of the process of data preprocessing of the present invention;

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

[0068] Figure 3 It is a three-dimensional time-frequency schematic diagram with time, frequency, energy and spatial distribution information generated by the present invention after data fusion;

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

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

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

[0072] The present invention is further described in detail below through specific examples. The following examples are only illustrative and not restrictive, and the protection scope of the present invention cannot be limited thereto.

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

[0074] S1. Data preprocessing: After completing the time-frequency conversion, frequency energy calculation and amplitude adjustment of the original IQ signal, a three-dimensional time-frequency diagram is obtained through data fusion;

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

[0076]

[0077] in:

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

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

[0080]

[0081] Where: Fs is the sampling frequency;

[0082] 2) Frequency point energy calculation: The input and output of the fast Fourier transform are both IQ complex signal data with a phase difference of 90 degrees, representing the real and imaginary parts of the complex signal respectively. The frequency point energy calculation performs modulo calculation on the real and imaginary parts 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), so the specific calculation formula of frequency point energy is:

[0084]

[0085] 3) Amplitude adjustment: Take the logarithm of the frequency energy to effectively observe signal changes. The calculation formula is:

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

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

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

[0089] x=[Pf n ] W,N =[Pf n ]W×N

[0090] Where: [Pf n ] W Indicates that the continuous frequency-power data Pf n Truncate according to the time window W to obtain the frequency-power data in each time window;

[0091] ×N means stacking the data of N time windows in the time dimension to form a three-dimensional time-frequency diagram;

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

[0093] like Figure 1 As shown in the figure, the original IQ signal data is converted into time-frequency, the frequency energy is calculated and the amplitude is adjusted to obtain the following Figure 2 The two-dimensional time-frequency diagram of different interferences shown in Figure 2 The two-dimensional time-frequency diagram objectively displays the spectrum characteristics of the interference signal that does not change with time at the current time node, such as Figure 2 b) random noise interference, c) broadband interference, and e) comb interference are clearly distinguished in the two-dimensional time-frequency diagram, but a) swept frequency interference and d) fixed frequency interference cannot be distinguished by the time-frequency diagram at a fixed moment.

[0094] Will Figure 2 The two-dimensional time-frequency data is generated by data fusion as shown in Figure 3 The three-dimensional time-frequency diagram with time, frequency, energy, and spatial distribution information is shown in Figure 3 The swept frequency interference characteristic shown is a slanted line that is discrete in time and has varying frequency intervals, while the fixed frequency interference is a continuous interference signal at a fixed frequency. The characteristics of the two are obvious.

[0095] S2, feature extraction of transfer learning: input the three-dimensional time-frequency graph into the pre-trained Resnet neural network for image feature extraction;

[0096] The feature extraction selects the 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 the Resnet neural network are Resnetout =Resnet(x); where Resnet out ∈R b×c′×h′×w′ .

[0097] S3, LSTM time series feature extraction: The obtained image features are transformed into dimensions and input into the LSTM network to obtain time series features;

[0098] 1) Resnet out The dimension conversion is performed according to the following formula:

[0099]

[0100] Where: T = h'×w';

[0101] 2) LSTM (input) Input into the LSTM network to obtain time series features:

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

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

[0104] S4: Input the image features obtained by S2 and the time series features obtained by S3 as input data to the Self-Attention Mechanism network, classify them through Linear and log_softmax, and complete the interference signal recognition;

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

[0106] 2) Interference type identification: It consists of Linear and log_softmax. The probability of each category is calculated through the log_softmax function to complete the interference signal identification.

[0107] like Figure 4 A schematic diagram of the network structure used in the present invention, wherein:

[0108] Feature extraction of transfer learning: The three-dimensional time-frequency graph generated in S1 is input 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 invention is not limited to residual network;

[0109] LSTM time series feature extraction: The feature data output by the Resnet neural network needs to be converted into a dimension through the View module, and then the time series feature extraction is performed through the LSTM module.

[0110] The image features obtained by S2 and the time series features obtained by S3 are input into the Self-Attention Mechanism network as input data, and classified by Linear and log_softmax to complete the interference signal recognition;

[0111] The Self-Attention Mechanism network consists of a multi-head attention module and an encoder module. The Embedding encoder module consists of a Linear layer, a Dropout layer, and a LaryerNorm layer.

[0112] Interference type identification: It consists of Linear and log_softmax. The probability of each category is calculated through the log_softmax function to complete the interference signal identification.

[0113] like Figure 5 , 6 This is the training convergence curve and interference recognition effect diagram of the present invention. It can be seen from the figure that:

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

[0115] (2) Figure 5 As shown in the figure, after 50 epochs of training, the loss of the training dataset and the validation dataset dropped below 0.1, and the model converged successfully.

[0116] (3) Figure 6The confusion matrix of the recognition results of different interferences on the test data set. Labels 0-5 represent six types of interference: no interference, swept frequency interference, random noise interference, broadband interference, comb interference, and fixed frequency interference. The recognition accuracy for no interference is 92%, which is relatively low mainly because there is spectrum interference from other equipment products in the space. The recognition accuracy in the presence of interference is above 98%.

[0117] (4) The present invention adopts the transfer learning method, which is very suitable for the situation of a small number of interference samples in actual application scenarios. Compared with the method without transfer learning, using 400 training data sets to train the feature extraction Resnet152 network in the network structure of the present invention cannot obtain a relatively good recognition accuracy.

[0118] Although the embodiments and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art will appreciate that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A communication interference identification method based on three-dimensional time-frequency and deep transfer learning, characterized by: The steps of the method are: S1. Data preprocessing: After completing the time-frequency conversion, frequency energy calculation and amplitude adjustment of the original IQ signal, a three-dimensional time-frequency diagram 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 series feature extraction: The obtained image features are transformed into dimensions and input into the LSTM network to obtain time series features; S4: Input the image features obtained in S2 and the time series features obtained in S3 into the self-attention mechanism network, assign weights through the attention module, and then combine Linear and log_softmax for classification to complete the interference signal recognition.

2. The communication interference identification method based on three-dimensional time-frequency and deep transfer learning according to claim 1 is characterized in that: The S1 is specifically: 1) Fast Fourier transform processing: The original IQ signal is processed by fast Fourier transform to complete the time-frequency conversion. For a vector s(n) containing N uniform sampling points, the fast Fourier transform formula of the original IQ signal is: in: k＝0～2 N -1； After fast Fourier transform, F n The signal frequency represented by any sampling point n is: Where: Fs is the sampling frequency; 2) Frequency point energy calculation: The input and output of the fast Fourier transform are both IQ complex signal data with a phase difference of 90 degrees, representing the real and imaginary parts of the complex signal respectively. The frequency point energy calculation performs modulo calculation on the real and imaginary parts of each frequency point signal to obtain the projection of the signal vector in the real number domain, that is, the signal energy. The data output after fast Fourier transform is Z(n)=a+bi(a,b∈R), so the specific calculation formula of frequency point energy is: 3) Amplitude adjustment: Take the logarithm of the frequency energy and the calculation formula is: P n =20log|z n |, 4) The frequency F corresponding to the sampling point n n and energy P n Forming frequency-power data: Pf n , 5) Generate three-dimensional time-frequency diagram: The continuous Pf n The data is truncated according to a fixed time window W, and then the data of N consecutive time windows are processed in three dimensions according to the time dimension to generate a three-dimensional time-frequency graph data x. The calculation formula is: x=[Pf n ] W,N =[Pf n ] W ×N Where: [Pf n ] W Indicates that the continuous frequency-power data Pf n Truncate according to the time window W to obtain the frequency-power data in each time window; ×N means stacking the data of N time windows in the time dimension to form a three-dimensional time-frequency diagram; The dimensions of x are: frequency f, power P, time t.

3. The communication interference identification method based on three-dimensional time-frequency and deep transfer learning according to claim 2 is characterized in that: Specifically, S2 is as follows: the feature extraction selects the Resnet neural network pre-trained on the ImageNet dataset, and the x(n) after data preprocessing 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 the Resnet neural network are Resnet out =Resnet(x); where Resnet out ∈R b×c×h×w , Resnet(x)=F(x,{W i })+x Where: F(x, {W i }) is the convolution operation in the residual block; W i is the weight parameter of the convolutional layer; +x represents the skip connection in the residual calculation.

4. The communication interference identification method based on three-dimensional time-frequency and deep transfer learning according to claim 3 is characterized in that: The S3 is specifically: 1) Resnet out The dimension conversion is performed according to the following formula: Where: T = h'×w' means that the data Resnet out ∈R b×c×h×w The h′×w′ two dimensions are spliced ​​together; 2) LSTM (input) Input into the LSTM(x) network to obtain the time series features. The calculation formula of LSTM(x) is as follows: i t =σ(w ii x t +b ii +w hi h t-1 +b hi ); f t =σ(w if x t +b if +w hf h t-1 +b hf ); g t =tanh(w ig x t +b ig +w hg h t-1 +b hg ); o t =σ(w io x t +b io +w ho h t-1 +b ho ); c t =f t ⊙c t-1 +i t ⊙g t ; h t =o t ⊙tanh(c t ); Where: i t 、f t , g t , o t They are input gate, forget gate, candidate unit state and output gate; c t is the unit state; h t is a hidden state; σ is the sigmoid activation function; ⊙ represents element-by-element multiplication; After the above formula is calculated, the output is: where d lstm is the dimension of the hidden layer of the LSTM network.

5. The communication interference identification method based on three-dimensional time-frequency and depth transfer learning according to claim 4 is characterized in that: The S4 is specifically: 1) Resnet out and LSTM (output) As input data SAM (input) To the self-attention mechanism network, the attention mechanism network consists of a multi-head attention module and an encoder module. The calculation process of the self-attention mechanism network is as follows: Q=MYSELF (input) IN Q ,K=SAM (input) IN K ,V=SAM (input) IN V ; SAM (output) =Embedding(Attention(Q,A,W)); Where: Q, K, V are query, key, and value matrices respectively; W Q , W K , W V is a learnable weight matrix; Embedding encoder module: consists of Linear layer, Dropout layer, and LaryerNorm layer; 2) Interference type identification: It consists of Linear and log_softmax. The probability of each category is calculated through the log_softmax function to complete the interference signal identification; The calculation formula of the interference signal identification network is: Y=log_softmax(linear(SAM out ,C)) Where: Y∈R b×c is the classification result; C is the number of categories.

Images