A pulse signal classification method based on intra-pulse fine features

Abstract

The application provides a pulse signal classification method based on intra-pulse fine features, and belongs to the field of electronic reconnaissance pulse signal classification and identification. The method comprises the following modules: intra-pulse fine feature extraction of the first type of pulse signal, intra-pulse fine feature extraction of the second type of pulse signal, intra-pulse fine feature extraction of the third type of pulse signal (which comprises the first type of pulse signal and the second type of pulse signal), long short-term memory (LSTM) training, and third type of pulse signal attribute prediction. Due to the differences in physical device attributes, the intra-pulse fine features of pulse signals generated by different devices are inevitably different. By analyzing the intra-pulse fine features of two types of pulse signals, comparing the two types of pulse signals by using LSTM, and giving a classification result, the method can avoid the false decision of our combat equipment on the battlefield situation. The method can effectively improve the correct classification probability of two / multiple types of pulse signals, greatly reduce the running time, and is simple to implement in engineering and has promotional value. The method can be applied to the classification of pulse signals emitted by different electronic reconnaissance devices, and can also be applied to the classification of daily wireless signals such as mobile phones and computers.

Description

Technical Field

[0001] This invention belongs to the field of electronic reconnaissance pulse signal classification and recognition and pulse signal processing, and relates to a pulse signal classification method, particularly a classification method for multiple types of pulse signals. Background Technology

[0002] In today's complex electronic warfare environment, various electronic reconnaissance devices continuously emit pulse signals. These different types of pulse signals superimpose in time and space, making it difficult for reconnaissance equipment to classify and identify useful signals, severely impacting its judgment of accurate information. To achieve accurate reception of various pulse signals and correct judgment of battlefield information in this environment, a core issue is the accurate and stable classification of different types of pulse signals.

[0003] To improve classification accuracy, previous engineering applications have commonly used classification methods based on pseudo-Wigner-Ville time-frequency images. However, this method involves a large amount of computation for processing two-dimensional images. In the field of electronic reconnaissance pulse signal classification and recognition, where real-time performance is critical, this method cannot guarantee real-time performance. Furthermore, this method is quite sensitive to signal parameters, and it requires significant adjustments as the duration of the transmitted pulse signal changes. Therefore, this method has not yet achieved significant practical engineering value.

[0004] Based on this, this paper proposes a pulse signal classification method based on subtle intrapulse features. Considering the characteristics of electronic reconnaissance pulse signal classification tasks, the existing selected features are adjusted. By extracting subtle intrapulse features, the real-time performance and classification accuracy can be effectively improved.

[0005] Purpose of the invention

[0006] The purpose of this invention is to propose a pulse signal classification method based on subtle intrapulse features, which solves the problems of large computational load, inability to guarantee timeliness, and high sensitivity to signal parameters in existing time-frequency image-based classification methods. Summary of the Invention

[0007] The objective of this invention is achieved as follows:

[0008] S1: Extraction of subtle features within the pulse of type 1 pulse signal;

[0009] S2: Extraction of subtle features within the pulse of type 2 pulse signal;

[0010] S3: LSTM training;

[0011] S4: Extraction of subtle features within the pulse of type 3 pulse signal;

[0012] S5: Type 3 pulse signal attribute prediction.

[0013] The present invention also includes the following structural features:

[0014] 1. Step S1 includes the following sub-steps:

[0015] S11: The received Type 1 pulse signal is sent to the analog-to-digital converter (ADC) module of the reconnaissance equipment receiver to convert the analog signal into a digital signal;

[0016] S12: Perform a Fast Fourier Transform (FFT) operation on the converted digital signal to obtain the discrete spectrum of the first type of pulse signal;

[0017] S13: Take the modulus of the discrete spectral components of the first type of pulse signal to obtain the modulus components;

[0018] S14: Use the peak search algorithm to search for the peak value of each modulus component of the first type of pulse signal, then divide each modulus component by the peak value to normalize the intrapulse subtle features of the first type of pulse signal, and extract the normalized intrapulse subtle features of the first type of pulse signal for training.

[0019] 2. Step S2 includes the following sub-steps:

[0020] S21: Send the Type 2 pulse signal to the ADC converter module of the reconnaissance equipment receiver to convert the analog signal into a digital signal;

[0021] S22: Perform FFT operation on the converted digital signal to obtain the discrete spectrum of the second type of pulse signal;

[0022] S23: Take the modulus of each discrete spectral component of the second type of pulse signal to obtain each modulus component;

[0023] S24: Use the peak search algorithm to search for the peak value of each modulus component of the second type of pulse signal, then divide each modulus component by the peak value to normalize the intra-pulse subtle features of the second type of pulse signal, and extract the normalized intra-pulse subtle features of the second type of pulse signal for training.

[0024] 3. The specific steps of step S3 are as follows:

[0025] S31: Training dataset preparation. According to the algorithm requirements, set the label of the first type of pulse signal to 1 and the label of the second type of pulse signal to 0.

[0026] S32: LSTM classifier construction;

[0027] S33: Read the training dataset into the LSTM and adjust the corresponding model parameters in a loop to achieve the best model performance;

[0028] 4. Step S4 includes the following sub-steps:

[0029] S41: The received Type 3 pulse signal with unknown attributes is sent to the ADC converter module of the reconnaissance equipment receiver to convert the analog signal into a digital signal;

[0030] S42: Perform FFT operation on the converted digital signal to obtain the discrete spectrum of the third type of pulse signal;

[0031] S43: Take the modulus of each discrete spectral component of the third type of pulse signal to obtain each modulus component;

[0032] S44: The peak value of each modulus component of the third type of pulse signal is searched using the peak value search algorithm. Then, each modulus component is divided by the peak value to normalize the intra-pulse subtle features of the third type of pulse signal and extract the normalized intra-pulse subtle features of the third type of pulse signal.

[0033] 5. Step S5 includes the following sub-steps:

[0034] S51: Normalize the intrapulse subtle features of the third type of pulse signal with unknown attributes from step S4;

[0035] S52: The LSTM classifier trained in step S3 can be used to predict the category of each pulse signal, which is either the first type or the second type.

[0036] Technical effect

[0037] Compared with existing technologies, the advantages of this invention are as follows: This invention only processes one-dimensional sequences, while the latter requires processing two-dimensional images. The computational load of this invention is much smaller than that of the background technology. Usually, to improve the correct classification probability, it is necessary to increase the number of training samples. As the number of training samples increases, the training time required by the background technology will increase exponentially. This invention has more outstanding timeliness performance, and both accuracy and timeliness performance can be well guaranteed. For changes in signal parameters, as long as the signal bandwidth remains unchanged, the classification effect of the method of this invention does not change much. However, the latter is more sensitive to changes in signal bandwidth and requires readjustment of network parameters to ensure good classification results. Therefore, the features extracted by this invention have better stability. Attached Figure Description

[0038] Figure 1 This is an overall flowchart of the pulse signal classification method based on subtle intrapulse features of the present invention;

[0039] Figure 2 This is a demonstration of the intrapulse subtle features of the first type of pulse signal in the pulse signal classification method based on intrapulse subtle features of the present invention;

[0040] Figure 3 This is a demonstration of the intrapulse subtle features of the second type of pulse signal in the pulse signal classification method based on intrapulse subtle features of the present invention. Detailed Implementation

[0041] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0042] like Figure 1 The diagram illustrates the pulse signal classification process of this invention. First, subtle intra-pulse features of the first type of pulse signal are extracted based on prior information for model training. Then, subtle intra-pulse features of the second type of pulse signal are extracted based on prior information for model training. Next, the network parameters are modified through LSTM iterative training to optimize network performance on the validation set. Then, subtle intra-pulse features of the third type of pulse signal with unknown attributes are extracted. Finally, the subtle intra-pulse features of the third type of pulse signal are input into the classification model, and the output result is obtained at the model's output, thus completing intelligent classification. The devices generating the first and second type of pulse signals are configured with a frequency domain intra-pulse fluctuation range of ±0.7 dB. The discrete spectrum sample vector lengths of the first and second type of pulse signals are 2048, the carrier frequency f0 = 2 GHz, the bandwidth B = 20 MHz, and the pulse width T = 20 μs. 1200 Monte Carlo simulations were performed on both Class 1 and Class 2 pulse signals under various signal-to-noise ratio conditions. 1000 simulation results were used to train the LSTM, and the remaining 200 simulation results were used to validate the classification results. Specific steps are detailed in the appendix. Figure 1 As shown.

[0043] A pulse signal classification method based on subtle intrapulse features according to the present invention is further implemented as follows:

[0044] S1: Extraction of subtle features within the pulse of type 1 pulse signal;

[0045] S11: Assume the system function in the frequency domain of the device generating the type 1 pulse signal is H1(f n ),have:

[0046] (1)H1(f n )=[Y1(f1),Y1(f2),…,Y1(f 2048 )];

[0047] Among them, Y1(f n ) represents a frequency of f n System function components of type 1 pulse signal transmitting equipment;

[0048] S12: Assume the spectrum of an ideal type I pulse signal is S1(f n),have:

[0049] (2)S1(f n )=[X1(f1),X1(f2),…,X1(f 2048 )];

[0050] Where X1(f n ) represents a frequency of f n The ideal spectral value for a Type I pulse signal;

[0051] S13: Due to equipment limitations, the normalized intra-pulse subtle features of the type 1 pulse signal |S′1(f n )| is:

[0052] (3)

[0053] Where '⊙' is the Hadamard product operator, S1(f n ) represents the offline spectrum of a type 1 pulse signal after being affected by the device, |S1(f n |S1(f) represents the subtle intrapulse features of the first type of pulse signal. max | is |S1(f n The peak search results of )| were used to perform 1000 Monte Carlo simulations according to Equation (3), thus completing the extraction of the normalized pulse subtle features of the first type of pulse signal used for training.

[0054] S2: Extraction of subtle features within the pulse of type 2 pulse signal;

[0055] S21: Assume the system function in the frequency domain of the device generating the type 2 pulse signal is H2(f n ),have:

[0056] (4)H2(f n )=[Y2(f1),Y2(f2),…,Y2(f 2048 )];

[0057] Among them, Y2(f n ) represents a frequency of f n The system function components of the second type of pulse signal transmitting equipment;

[0058] S22: Assume the spectrum of an ideal type 2 pulse signal is S2(f n ),have:

[0059] (5)S2(f n )=[X2(f1),X2(f2),…,X2(f 2048 )];

[0060] S23: Due to equipment limitations, the normalized intra-pulse subtle features of the type 2 pulse signal |S′2(f n )| is:

[0061] (6)

[0062] Where '⊙' is the Hadamard product operator, S2(f n ) represents the discrete spectrum of a type 2 pulse signal after being affected by the device, |S2(f n |S2(f) represents the subtle intrapulse characteristics of the second type of pulse signal. max | is |S2(f n The peak search results of )| were used to perform 1000 Monte Carlo simulations according to Equation (6), thus completing the extraction of the normalized pulse subtle features of the second type of pulse signal used for training.

[0063] S3: LSTM classifier training;

[0064] S31: Preprocess the training dataset generated by the Monte Carlo simulation, setting the labels of 1000 known Class 1 pulse signals to 1 and the labels of 1000 known Class 2 pulse signals to 0;

[0065] S32: LSTM classifier preparation, set the number of LSTM feature extraction layers to 3, the number of fully connected layers to 2, the input feature dimension to 2048, and the hidden layer feature dimension to 256;

[0066] S33: Build and train the network in the Jupyter Notebook compilation environment;

[0067] S4: Extraction of subtle features within the pulse of type 3 pulse signal;

[0068] S41: Assume the system function of the domain of the device generating the type 3 pulse signal is H3(f n ), which is H1(f n ) or H2(f n ),have:

[0069] (7)H3(f n )=[Y3(f1), Y3(f2),…,Y3(f 2048 )];

[0070] Among them, Y3(f n ) represents a frequency of f n The system function components of the third type of pulse signal transmitting equipment;

[0071] S42: Assume the spectrum of an ideal type 3 pulse signal is S3(f n ),have

[0072] (8)S3(f n )=[X3(f1),X3(f2),…,X3(f 2048 )];

[0073] S43: Due to equipment limitations, the normalized intra-pulse subtle features of the type 3 pulse signal |S′3(f n )| is:

[0074] (9)

[0075] Where '⊙' is the Hadamard product operator, S3(f n ) represents the discrete spectrum of a type 3 pulse signal after being affected by the device, |S3(f n | represents the subtle intrapulse characteristics of the second type of pulse signal, |S3(f max | is |S3(f n The peak search results of )| were used to perform 1000 Monte Carlo simulations according to Equation (9), thus completing the extraction of the normalized pulse subtle features of the second type of pulse signal used for training.

[0076] S5: Pulse signal attribute prediction;

[0077] S51: Normalize the 400 received Type 3 unknown attribute pulse signals to be predicted in step S4 to obtain intrapulse subtle features |S′3(f) n | Read in;

[0078] S52: Call the LSTM prediction program to predict the category of each pulse signal, whether it belongs to the first or second type of pulse signal.

[0079] The classification accuracy of this embodiment is shown in Table 1. It can be seen that the pulse signal classification method of this embodiment has a high classification accuracy. At the same time, the time required by this embodiment is 3 to 6 seconds, while the time required by the method using two-dimensional time-frequency images under the same conditions is 10 to 20 times that of this embodiment. The method of this embodiment can greatly reduce the time.

[0080] Table 1. Classification accuracy (%) of the pulse signal classification method of the present invention on the test set.

[0081]

[0082]

[0083] The above description is merely one application example of the present invention, intended only to help those skilled in the art better understand the specific implementation ideas of the present invention, and is not intended to limit the present invention in any way. Any modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A pulse signal classification method based on subtle intrapulse features, characterized in that... Includes the following steps: S1: Extraction of subtle features within the pulse of type 1 pulse signal; S11: Assume the system function in the frequency domain of the device generating the type 1 pulse signal is H1(f n ),have: (1)H1(f n )=[Y1(f1), Y1(f2),…, Y1(f 2048 )]; Among them, Y1(f n ) represents a frequency of f n System function components of type 1 pulse signal transmitting equipment; S12: Assume the spectrum of an ideal type I pulse signal is S1(f n ),have: (2)S1(f n )=[X1(f1),X1(f2),…,X1(f 2048 )]; Where X1(f n ) represents a frequency of f n The ideal spectral value for a Type I pulse signal; S13: Due to equipment limitations, the normalized intra-pulse subtle features of the type 1 pulse signal |S′1(f n )| is: (3) Where '⊙' is the Hadamard product operator, S1(f n ) represents the offline spectrum of a type 1 pulse signal after being affected by the device, |S1(f n |S1(f) represents the subtle intrapulse features of the first type of pulse signal. max | is |S1(f n The peak search results of )| were used to perform 1000 Monte Carlo simulations according to Equation (3), thus completing the extraction of the normalized pulse subtle features of the first type of pulse signal used for training. S2: Extraction of subtle features within the pulse of type 2 pulse signal; S21: Assume the system function in the frequency domain of the device generating the type 2 pulse signal is H2(f n )),have: (4)H2(f n )=[Y2(f1),Y2(f2),…,Y2(f 2048 )]; Among them, Y2(f n ) represents a frequency of f n The system function components of the second type of pulse signal transmitting equipment; S22: Assume the spectrum of an ideal type 2 pulse signal is S2(f n ),have: (5)S2(f n )=[X2(f1),X2(f2),…,X2(f 2048 )]; S23: Due to equipment limitations, the normalized intra-pulse subtle features of the type 2 pulse signal |S′2(f n )| is: (6) Where '⊙' is the Hadamard product operator, S2(f n ) represents the discrete spectrum of a type 2 pulse signal after being affected by the device, |S2(f n |S2(f) represents the subtle intrapulse characteristics of the second type of pulse signal. max | is |S2(f n The peak search results of )| were used to perform 1000 Monte Carlo simulations according to Equation (6), thus completing the extraction of the normalized pulse subtle features of the second type of pulse signal used for training. S3: LSTM classifier training; S31: Preprocess the training dataset generated by the Monte Carlo simulation, setting the labels of 1000 known Class 1 pulse signals to 1 and the labels of 1000 known Class 2 pulse signals to 0; S32: LSTM classifier preparation, set the number of LSTM feature extraction layers to 3, the number of fully connected layers to 2, the input feature dimension to 2048, and the hidden layer feature dimension to 256; S33: Build and train the network in the Jupyter Notebook compilation environment; S4: Extraction of subtle features within the pulse of type 3 pulse signal; S41: Assume the system function of the domain of the device generating the type 3 pulse signal is H3(f n ), which is H1(f n ) or H2(f n ),have: (7)H3(f n )=[Y3(f1),Y3(f2),…,Y3(f 2048 )]; Among them, Y3(f n ) represents a frequency of f n The system function components of the third type of pulse signal transmitting equipment; S42: Assuming the spectrum of an ideal type 3 pulse signal is S3(fn), we have (8)S3(f n )=[X3(f1),X3(f2),…,X3(f) 2048 )]; S43: Due to equipment limitations, the normalized intra-pulse subtle features of the type 3 pulse signal |S′3(f n )| is: (9) Where '⊙' is the Hadamard product operator, S3(f n ) represents the discrete spectrum of a type 3 pulse signal after being affected by the device, |S3(f n | represents the subtle intrapulse characteristics of the second type of pulse signal, |S3(f max | is |S3(f n The peak search results of )| were used to perform 1000 Monte Carlo simulations according to Equation (9), thus completing the extraction of the normalized pulse subtle features of the second type of pulse signal used for training. S5: Pulse signal attribute prediction; S51: Normalize the 400 received Type 3 unknown attribute pulse signals to be predicted in step S4 to obtain intrapulse subtle features |S3(f n | Read in; S52: Call the LSTM prediction program to predict the category of each pulse signal, whether it belongs to the first or second type of pulse signal.

Images