Lfmc feature and transformer integrated classifier based lfm signal classification method

Abstract

The application discloses an LFM signal classification method based on an MFCC feature and a Transformer integrated classifier, carries out signal pretreatment on collected linear frequency modulation (LFM) signals to obtain effective pulse signals, carries out an MFCC feature extraction process including pre-emphasis, framing, windowing, fast Fourier transformation, a Mel filter group, logarithmic operation, discrete cosine transformation and dynamic differential parameter extraction to obtain static basic features, first-order differential dynamic features and second-order differential dynamic features; the three groups of features are respectively input into three Transformer classifiers for differential training to obtain pre-classification features; the three groups of pre-classification features are combined and input into an integrated module, are respectively subjected to normalization and three-layer linear layer operation, and finally, a final classification result is output through a full connection layer. The application combines the MFCC feature and the Transformer, proposes a Transformer integrated classification method suitable for the LFM signal, and effectively solves the problem that it is difficult to distinguish the same signal source.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of information security, and mainly relates to an LFM signal classification method based on an MFCC feature and a Transformer integrated classifier. BACKGROUND

[0002] Spread spectrum communication, optical fiber communication and satellite communication are known as the three major communication transmission methods in the information age. Among them, the development time is the longest, and the frequency band expansion of spread spectrum communication is generally completed through pseudo-random code to achieve anti-interference and multi-address communication. Linear frequency modulation is a spread spectrum modulation technology without pseudo-random code sequence. Linear frequency modulation signal has longer duration than the same bandwidth signal, so it is widely used in radar, sonar and other scenes. Through the linear frequency modulation signal emitted from the radar, sonar and other devices, it is very important to identify the specific transmitter of the signal. Therefore, the industry needs an efficient and high-accuracy LFM signal source identification method.

[0003] Most of the traditional identification methods are based on pulse description word technology, which extracts the frequency, arrival time and other shallow features of the pulse signal for identification. However, with the increasing complexity of the electromagnetic environment and the continuous improvement of electronic technology, when the signal source of the linear frequency modulation signal has the same configuration and parameter setting, the identification work is often difficult to carry out. Because the extracted features are extremely similar, the classifier often cannot work well, and the radio frequency fingerprint technology can well solve this problem. Each signal source transmitter will have some hardware differences when manufactured. These differences caused by hardware defects are the inherent hardware characteristics of the transmitter, which are often unique, like human fingerprints, so these characteristics are called "radio frequency fingerprints". Through in-depth research, radio frequency fingerprints are considered to be the most ideal features of wireless devices. Even if the frequency and modulation method of the transmitter are the same, the radio frequency fingerprint features can also work well. Radio frequency fingerprint features can be roughly divided into frequency offset, time domain envelope, modulation domain features, spectral features, etc. The most commonly used is the spectral feature extracted by Mel-frequency Cepstral Coefficients (MFCC). Therefore, MFCC is a potentially feasible method for solving the difficulty of LFM signal identification.

[0004] However, to efficiently identify the signal source, only the characteristics are far from enough, and a good classifier also plays a very key role. Since the Transformer was proposed, it has become the most popular model in the NLP field, but the advantage of the model is not only its excellent classification ability, but more importantly, it can be applied to many fields and has strong versatility, such as the Vision Transformer model used in the CV field. Therefore, it is a feasible scheme to combine MFCC with the Transformer. SUMMARY

[0005] The present application is aimed at the problem of difficulty in identifying the same configuration of the linear frequency modulation signal transmitter in the prior art, and provides a LFM signal classification method based on the MFCC feature and the Transformer integrated classifier. The collected linear frequency modulation signal is subjected to signal preprocessing to obtain an effective pulse signal. Then, for each effective pulse signal, an MFCC feature extraction process including pre-emphasis, framing, windowing, fast Fourier transform, Mel filter bank, logarithmic operation, discrete cosine transform, and dynamic difference parameter extraction is performed to obtain static basic features, first-order difference dynamic features, and second-order difference dynamic features. The three groups of features are respectively input into three Transformer classifiers for differential training to obtain pre-classification features. Then, the three groups of pre-classification features obtained by the three Transformer classifiers are combined and input into an integrated module for normalization and three-layer linear layer operation, and finally, a final classification result is output through a full connection layer. The present application combines the MFCC feature and the Transformer, and proposes a Transformer integrated classification method suitable for the linear frequency modulation signal, which effectively solves the problem of difficulty in distinguishing the same configuration of the signal source.

[0006] To achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows: a LFM signal classification method based on the MFCC feature and the Transformer integrated classifier, comprising the following steps:

[0007] S1, signal preprocessing: the collected linear frequency modulation signal is subjected to signal preprocessing to obtain an effective pulse signal, and the preprocessing process at least includes framing processing and endpoint determination;

[0008] S2, MFCC feature extraction: for each effective pulse signal obtained in step S1, the MFCC feature extraction is performed to obtain static basic features, first-order difference dynamic features, and second-order difference dynamic features, and the feature extraction process includes pre-emphasis, framing, windowing, fast Fourier transform, Mel filter bank, logarithmic operation, discrete cosine transform, and dynamic difference parameter extraction;

[0009] S3, classifier training: input three groups of features into three Transformer classifiers respectively for differential training, the classifier at least includes two modules: a segmentation module and a classification module; in the segmentation module, the input feature data is normalized to generate a new feature map suitable for the Transformer classifier, the new feature map is segmented to obtain a sub-feature map, and the sub-feature map is linearly mapped, so that the initial two-dimensional feature map is converted into a one-dimensional vector; after passing through the classification module, a pre-classification feature is obtained;

[0010] S4, data integration: after merging the three groups of pre-classification features obtained by the three Transformer classifiers in step S3, input into the integration module, respectively, normalize and three linear layers, and finally output the final classification result through a fully connected layer.

[0011] As an improvement of the application, the frame length and frame shift contained in each frame in the frame division process in step S1 are determined according to the sampling rate; the starting frame and the ending frame of each pulse signal are determined by calculating the product of the short-time zero-crossing rate and the short-time energy, and the starting point is the first time when the product is greater than the threshold value, and the ending point is the first time when the product is less than the threshold value.

[0012] As another improvement of the application, in the frame division process in step S1, the frame shift is half of the frame length; in the filter bank stage in step S2, the number of filter banks M is 22-26; the MFCC coefficient obtained by the discrete cosine transform stage is half, that is, the MFCC coefficient dimension is

[0013] As another improvement of the application, in the segmentation module of step S3, the MFCC feature map with a dimension of LxD is divided into N non-overlapping sub-feature maps in order, and the size of each sub-feature map is lxd, wherein Then linearly map each sub-feature map, so that the initially given LxD two-dimensional feature map is converted into N lxd one-dimensional vectors Wherein d takes a value of The value of l is the same as that of d.

[0014] As another improvement of the application, in step S4, the three linear layer nodes take values of 128, 64 and 32 respectively.

[0015] Compared with the prior art, the method has the beneficial effects that: a LFM signal classification method based on MFCC features and a Transformer integrated classifier is provided, and the method can achieve a very ideal accuracy rate in the identification task of multiple same configuration transmitters. The method converts the linear frequency modulation signal into three groups of MFCC features that can reflect the differences of each transmitter through a series of signal processing methods such as FFT and DCT, obtains a pre-classification result through a Transformer classifier, and then comprehensively processes three difference training results through an integrated module to obtain the final classification result. The method effectively solves the problem of classification difficulty of same configuration transmitters. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A step flow chart of the method of the present application;

[0017] Figure 2 A step flow chart of the MFCC feature extractor in step S2 of the method of the present application;

[0018] Figure 3 A step flow chart of the Transformer feature extractor in step S3 of the method of the present application;

[0019] Figure 4 An architecture diagram of the multi-layer perceptron in step S3 of the method of the present application;

[0020] Figure 5 An architecture diagram of the integrated module in step S4 of the method of the present application;

[0021] Figure 6 A ROC curve comparison diagram of the ordinary Transformer classifier and the Transformer integrated classifier in the test example of the present application for distinguishing the second signal source;

[0022] Figure 7 A ROC curve comparison diagram of the ordinary Transformer classifier and the Transformer integrated classifier in the test example of the present application for distinguishing the sixth signal source. DETAILED DESCRIPTION

[0023] The present application will be further illustrated below in combination with the drawings and specific embodiments, and it should be understood that the following specific embodiments are only used to illustrate the present application and not to limit the scope of the present application.

[0024] Embodiment 1

[0025] A LFM signal classification method based on MFCC features and a Transformer integrated classifier, through which same configuration signal source transmitters can be accurately and efficiently classified. The step flow of the method is as shown in Figure 1 .

[0026] Step S1: Preprocessing of the signal. Since the collected linear frequency modulation signal is a whole signal, preprocessing of frame division and endpoint determination is required.

[0027] 1. Frame division stage: according to the sampling rate, determine the sampling points contained in each frame, i.e. frame length and frame shift. The collected signal is an LFM modulated pulse signal, and the receiver sampling rate should be greater than the Nyquist sampling rate, i.e. oversampling is required. According to the sampling rate, the frame length and frame shift are determined, and the frame shift is usually half of the frame length.

[0028] 2. Endpoint determination stage: determine the starting frame and ending frame of each pulse signal by calculating the product of short-time zero-crossing rate and short-time energy. The starting point of each valid pulse signal is the position where the product of short-time zero-crossing rate and short-time energy first exceeds the threshold value, and the ending point of each valid pulse signal is the position where the product of short-time zero-crossing rate and short-time energy first falls below the threshold value. Through these two steps of processing, the valid pulse signal s[i] is obtained.

[0029] Step S2: For each valid pulse signal, extract the MFCC feature, which includes eight stages: pre-emphasis, frame division, windowing, Fast Fourier Transform (FFT), Mel filter bank, logarithmic operation, Discrete Cosine Transform (DCT), and dynamic difference parameter extraction, as shown in Figure 2 .

[0030] 1. Pre-emphasis stage: each signal passes through a high-pass filter to make the overall signal spectrum flat, and the formula is as follows:

[0031] s[i] = s[i+1] - k * s[i]

[0032] where k is a self-defined parameter, and the recommended value is 0.95.

[0033] 2. Frame division stage: this stage divides each valid pulse signal into N frames of the same size, similar to the frame division in step one, and x(n) represents the nth frame.

[0034] 3. Windowing stage: use the Hamming window h to process each frame of signal to increase the continuity of the left and right segments, and the formula is as follows:

[0035]

[0036] 4. FFT stage: perform Fast Fourier Transform on the signal to obtain the energy spectrum F(k), where K is the number of Fourier Transform points, and the formula is as follows:

[0037]

[0038] 5. Mel filter bank stage: This stage smoothes the energy spectrum by passing it through a bank of triangular filters on a Mel scale, removing the effect of harmonics. The frequency response of the triangular filter is given by:

[0039]

[0040] where f() is the Mel frequency, m = 1, 2, …, M, and M is the number of filters.

[0041] The dimension of the MFCC coefficients depends on the number of Mel filters M in step 2, which is usually taken as 22-26.

[0042] 6. Logarithm operation stage: Calculate the logarithmic energy of each filter bank output, given by:

[0043]

[0044] 7. DCT stage: Obtain MFCC coefficients through DCT, given by:

[0045]

[0046] Through the DCT stage, M-dimensional MFCC coefficients are obtained, usually only the first half is retained, so the dimension of the MFCC coefficients is

[0047] 8. Dynamic difference parameter extraction stage: The MFCC coefficients obtained through the DCT stage are only static features, by calculating the first-order difference parameters and the second-order difference parameters, dynamic features are obtained, given by:

[0048]

[0049] Through the above eight stages, three groups of MFCC features are obtained as the final input into the Transformer classifier.

[0050] Step S3: For each group of features, input into three Transformer classifiers for differential training. The classifier can be roughly divided into two modules: segmentation module and classification module, as shown in Figure 3 In the segmentation module, first, the input feature data is normalized to generate a new feature map suitable for the Transformer classifier. Then the feature map is segmented, and the MFCC feature map with dimension LxD is divided into N non-overlapping sub-feature maps in order, each with size lxd, where Then each sub-feature map is linearly mapped, so that the initially given two-dimensional feature map of LxD is converted into N one-dimensional vectors of lxd

[0051] where the size of the sub-feature map has a large impact on the result, and it is usually recommended that d take a value of l takes the same value as d.

[0052] Normalization is performed on each sub-feature map. The formula is as follows:

[0053]

[0054] where x i is the sub-feature map, and a and b are learnable parameters, μ L is the mean of the sample, σ is the variance of the sample, and e is a very small threshold value to prevent errors caused by the denominator being 0.

[0055] The attention weight of each sub-feature map is calculated, and the formula is as follows:

[0056]

[0057] where Q, K, and V are obtained by multiplying the sub-feature map with the coefficient matrix, and D k represents the dimension of K. is a scaling factor to prevent adverse effects on backpropagation when the dimension is very large.

[0058] Q, K, and V are divided into h parts through linear mapping, denoted as head1, head2, …, head h , and the attention weight of each part is calculated. Finally, the attention scores in the h subspaces are combined, and a linear mapping is performed to obtain the final output. The formula of the multi-head self-attention mechanism is as follows:

[0059] Multihead(Q, K, V) = Concat(head1, head2, …, head h )W 0

[0060] where Concat means merging head1, head2, …, head h , and W 0 is a coefficient matrix.

[0061] The attention weight vector obtained in the above steps is normalized.

[0062] The attention weight vector is processed by a multi-layer perceptron, such as Figure 4As shown, the attention weight vector of the entire feature map is obtained by sequentially passing through fully connected layer A, activation layer, fully connected layer B, and dropout layer. The role of the dropout layer is to prevent overfitting during training, and its value is usually set to 0.1-0.3.

[0063] The above operations, from normalizing each sub-feature map to this point, need to be repeated N times, where N is usually 6-8.

[0064] The weight vector is normalized, flattened, and discarded, and then processed again by a multilayer perceptron. The activation layer in the multilayer perceptron can be GELU, RULE, or Sigmoid.

[0065] The pre-classification features of each pulse signal in the classifier are obtained by passing through a fully connected layer.

[0066] Step S4: Merge the three sets of pre-classification features obtained from the three Transformer classifiers and input them into the integration module, such as... Figure 5 As shown, normalization and three linear layer operations are performed respectively, and finally the classification result is obtained through a fully connected layer. The values ​​of the nodes in the three linear layers have a great influence on the result, and they are usually set to 128, 64 and 32 respectively.

[0067] Test case

[0068] This test case uses simulated signals to verify the performance of the proposed method. First, each simulated signal is preprocessed to obtain an effective pulse signal. Then, the effective pulse signal is input into the MFCC feature extractor to obtain three sets of feature maps. Next, each set of feature maps is input into a different Transformer classifier for individual training. Finally, the three training results are input into the integration module to obtain the final classification result.

[0069] This test case studies 3000 samples from six signal sources, including 2100 samples in the training set and 900 samples in the test set, achieving a final classification accuracy of 96%. This invention does not simply combine MFCC with a Transformer; instead, it divides the features of MFCC into three groups and inputs them into different Transformers for differential training, obtaining the final result through an ensemble module. Analysis of the results reveals that the classifier is not sensitive to the features of the second and sixth signal sources. Therefore, the prediction results from these two sources are compared in a comparative experiment. The experiment demonstrates that the proposed ensemble classifier method achieves better results. Figure 6 This is a comparison chart of the second signal source. Figure 7 This is a comparison chart for the sixth signal source. Method 1 in the chart refers to the single classifier method, and Method 2 refers to the ensemble classifier method. From... Figure 6 and Figure 7It can be seen that the ROC curve of method two deviates more from the 45-degree diagonal, indicating that the sensitivity and specificity reach a higher level in the case of mutual balance. In addition, the higher AUC value implies that the classifier model has higher prediction value under proper threshold setting.

[0070] It should be noted that the above only illustrates the technical idea of the present application, and cannot limit the protection scope of the present application. For those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which fall within the scope of protection of the claims of the present application.

Claims

1. An LFM signal classification method based on MFCC features and a Transformer ensemble classifier, characterized in that, Includes the following steps: S1, Signal preprocessing: The acquired linear frequency modulated signal is preprocessed to obtain an effective pulse signal. The preprocessing process includes at least frame segmentation and endpoint determination. S2, MFCC feature extraction: For each valid pulse signal obtained in step S1, MFCC feature extraction is performed to obtain static basic features, first-order differential dynamic features and second-order differential dynamic features. The feature extraction process includes pre-emphasis, framing, windowing, fast Fourier transform, Mel filter bank, logarithmic operation, discrete cosine transform and dynamic differential parameter extraction. S3, Classifier Training: The three sets of features are input into three Transformer classifiers for differential training. Each classifier includes at least two modules: a segmentation module and a classification module. In the segmentation module, the input feature data is normalized to generate a new feature map suitable for the Transformer classifier. The new feature map is segmented to obtain sub-feature maps, and then the sub-feature maps are linearly mapped to transform the initial two-dimensional feature map into a one-dimensional vector. After passing through the classification module, pre-classified features are obtained. S4, Data Integration: The three sets of pre-classification features obtained from the three Transformer classifiers in step S3 are merged and input into the integration module. Normalization and three linear layer operations are performed respectively, and finally, the final classification result is obtained through a fully connected layer.

2. The LFM signal classification method based on MFCC features and a Transformer ensemble classifier as described in claim 1, characterized in that: The framing process in step S1 involves determining the frame length and frame shift of each frame based on the sampling rate. The endpoints are determined by calculating the product of the short-time zero-crossing rate and the short-time energy to determine the start and end frames of each pulse signal. The first time the product is greater than a threshold within a relative time period is the start point, and the first time the product is less than a threshold is the end point.

3. The LFM signal classification method based on MFCC features and a Transformer ensemble classifier as described in claim 2, characterized in that: In step S2 Pre-emphasis stage: Each effective pulse signal passes through a high-pass filter to flatten the overall signal spectrum, as shown in the following formula: s[i] = s[i+1] - k*s[i] Where s[i] is the effective pulse signal; k is a user-defined parameter; Framing stage: Each valid pulse signal is divided into N frames of the same size, the same as the framing in step S1, and x(n) is denoted as the nth frame; Windowing stage: A Hamming window h is used to process each frame of signal to increase the continuity between the left and right segments, as shown in the following formula:

4. The LFM signal classification method based on MFCC features and a Transformer ensemble classifier as described in claim 3, characterized in that: In step S2 Fast Fourier Transform (FFT) stage: The signal is processed by Fast Fourier Transform (FFT) to obtain the energy spectrum F(k). Let K be the number of points in the Fourier Transform, and the formula is as follows: Mel filter bank stage: The energy spectrum is smoothed by passing it through a set of Mel-scale triangular filters to eliminate harmonic effects. The frequency response formula of the triangular filter is as follows: Where f() is the Mel frequency, m=1,2,…,M, and M is the number of filters; Logarithmic operation stage: Calculate the logarithmic energy of the output of each filter bank, using the following formula:

5. The LFM signal classification method based on MFCC features and a Transformer ensemble classifier as described in claim 3 or 4, characterized in that: In step S2 Discrete Cosine Transform Stage: The MFCC coefficients are obtained through the discrete cosine transform, as shown in the following formula: Dynamic difference parameter extraction stage: The MFCC coefficients obtained after the discrete cosine transform stage are only static features. By calculating the first-order and second-order difference parameters, the dynamic features are obtained, as shown in the following formula:

6. The LFM signal classification method based on MFCC features and a Transformer ensemble classifier as described in claim 5, characterized in that: In the framing process of step S1, the frame shift is half the frame length; in the Mel filter bank stage of step S2, the number of filter banks M is 22-26; the MFCC coefficients obtained by the discrete cosine transform stage are half the value, that is, the dimension of the MFCC coefficients is...

7. The LFM signal classification method based on MFCC features and a Transformer ensemble classifier as described in claim 5, characterized in that: The work of the classification module in step S3 includes: S31: Perform a normalization operation on each sub-feature map, using the following formula: Where, x i Sub-feature maps; α and β are learnable parameters; μ L The mean of the sample; denoted as the sample variance; ∈ represents the threshold. S32: Calculate the attention weights for each sub-feature map, using the following formula: Where Q, K, and V are all obtained by multiplying the sub-feature map with the coefficient matrix; D k Indicates the dimension of K; S33: Divide Q, K, and V into h parts using a linear mapping, denoted as head1, head2, ..., head... h The attention weights in step S32 are calculated for each sample, and the attention scores in the h-subspace are merged and linearly mapped to obtain the final output. The formula for the multi-head self-attention mechanism is as follows: Multihead(Q,K,V)=Concat(head1,head2,...,head h )W 0 Wherein, Concat refers to concatenating head1, head2, ..., head... h Merge; W 0 It is a coefficient matrix; S34: Perform the normalization operation in step S31 on the attention weight vector obtained in step S33; S35: The attention weight vector is processed by a multilayer perceptron, that is, it passes through a fully connected layer A, an activation layer, a fully connected layer B and a dropout layer in sequence to obtain the attention weight vector of the entire feature map. S36: Normalize, flatten, and discard the weight vector, and then process it again through the perceptron in step S35. S37: The pre-classification features of each pulse signal in the classifier are obtained through a fully connected layer.

8. The LFM signal classification method based on MFCC features and a Transformer ensemble classifier as described in claim 7, characterized in that: In the segmentation module of step S3, the L×D MFCC feature map is sequentially divided into N non-overlapping sub-feature maps, each sub-feature map having a size of l×d. Next, a linear mapping is performed on each sub-feature map, transforming the initially given L×D two-dimensional feature map into N one-dimensional vectors of l×d. Where d takes the value of The value of l is the same as that of d.

9. The LFM signal classification method based on MFCC features and a Transformer ensemble classifier as described in claim 8, characterized in that: In step S4, the values ​​of the nodes in the three linear layers are 128, 64, and 32, respectively.

Images