Method for Inter-Pulse Modulation Mode Recognition of Radar Emitter Based on TCN Network

Abstract

A method for identifying the inter-pulse modulation mode of radar radiation sources based on a TCN network, belonging to the technical field of signal processing. The purpose of the present invention is to input the radar radiation source pulse signal sequence into the TCN network for feature extraction and recognition based on convolutional operations, so as to improve the recognition rate of radar pulse signals. The steps of the present invention include generating training set and test set data respectively; preprocessing the generated data set signals; constructing a TCN network and setting the parameters of the TCN network; inputting the training set and test set data into the TCN network and further processing the data according to the algorithm requirements; iterating the network in the "training-test" manner, and when iterating to the nth time, ending the training, and the network output is the predicted category of the radar radiation source modulation mode. The present invention inputs the time-domain and frequency-domain signals after denoising the radar radiation source pulse signals into the TCN network respectively for feature extraction and recognition based on convolutional operations, thereby realizing the improvement of the recognition rate of radar pulse signals.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing. Background Art

[0002] With the development of electronic and communication technologies, radar systems play a crucial role in modern warfare. Facing the increasingly fierce form of electronic countermeasures and the continuous development of radar system-related technologies, radar signal recognition, as a key technology, has become a decisive factor in electronic warfare.

[0003] With the rapid development of artificial intelligence (AI) technology, deep learning (DL) technology has been widely applied in fields such as natural language processing and computer vision. A deep learning model is a machine learning model based on an artificial neural network. Through the mutual connection and calculation of multi-level neurons, it learns and extracts feature information from data, and completes tasks such as data classification, clustering, and regression. Currently, the neural networks for radar pulse signal recognition are mainly networks such as RNN. However, for long-time sparse radar pulse signals with missing and false pulses, the above networks are extremely prone to problems such as gradient explosion or gradient disappearance during operation, resulting in network failure and low classification accuracy. Therefore, it is necessary to design a recognition method for the inter-pulse modulation mode of long-time sparse radar radiation sources with missing and false interferences to solve the above technical problems. Summary of the Invention

[0004] The object of the present invention is a recognition method for the inter-pulse modulation mode of radar radiation sources based on a TCN network, which inputs a radar radiation source pulse signal sequence into the TCN network for feature extraction recognition based on convolutional operations, thereby improving the recognition rate of radar pulse signals.

[0005] The steps of the present invention are as follows:

[0006] S1. Preprocess the training set and test set signals:

[0007] S01. Perform a discrete Fourier transform on the radar pulse sequence with interference data to obtain a frequency-domain signal with noise. Set the points less than 10% of the maximum amplitude to 0 in the frequency spectrum to achieve the effect of noise reduction and obtain a denoised frequency-domain signal;

[0008] S02. Perform an inverse Fourier transform on the denoised frequency-domain signal, using 0.5 as the amplitude reference, to obtain a denoised binary time-domain signal;

[0009] S2. Construct a TCN network and set the TCN network parameters:

[0010] S01. Set the number of output channels of the dilated causal convolution to [16, 32, 32, 32, 16], set the number of network layers to 5, set dropout to 0.1, set the kernel size to 5, set the loss function of the TCN network to the negative log-likelihood loss function, use the AdamOptimizer algorithm to control the learning rate, use the rectified linear unit function as the activation function, and use the LogSoftmax classifier as the output layer of the network;

[0011] S02. The neurons in the TCN network are represented as follows:

[0012] (1)

[0013] In the formula represents the filter; represents the sequence; d represents the dilation factor; t represents time; k represents the kernel size;

[0014] S03. The residual block in the TCN network is represented as follows:

[0015] (2)

[0016] In the formula represents the activation function; represents the input sequence; represents the sequence after the input sequence passes through the convolution operation;

[0017] S04. The rectified linear unit activation function is represented as follows:

[0018] (3)

[0019] In the formula, represents the input value, which is a constant coefficient, represents the output of the rectified linear unit activation function;

[0020] S05. The LogSoftmax classifier is represented as follows:

[0021] (4)

[0022] In the formula, represents the i-th element in the data, represents taking the exponential of the i-th element, represents taking the sum of the exponentials of all elements, represents the scoring value output for the i-th element.

[0023] The TCN network adopted by the present invention takes as input a long - time sparse radar radiation source pulse signal sequence with missing and false interferences, reduces the influence of interference data by denoising in the frequency domain. After the denoised time - domain and frequency - domain signals are respectively extracted with features by the TCN network, the feature vectors of the two are spliced and fused as the recognition basis, thereby realizing the efficient and accurate recognition of signals with different modulation methods and avoiding problems such as gradient explosion. Therefore, the present invention has important practical application value in the field of radar signal processing technology. Brief Description of the Drawings

[0024] Figure 1 is the implementation flowchart of the present invention;

[0025] Figure 2 is the recognition accuracy of the network under different working conditions. Detailed Embodiments

[0026] In order to enable those skilled in the art to better understand the technical solution of the present invention, the preferred implementation embodiments of the present invention are described below with specific examples.

[0027] As Figure 1 shown, a method for identifying the inter - pulse modulation mode of a radar radiation source based on a TCN network, the identification method includes the following steps:

[0028] S1. Simulate and generate five typical radar pulse signals, including conventional pulse signals, staggered pulse signals, group - variable and agile pulse signals, sliding - variable pulse signals, and jittering pulse signals. The pulse repetition interval (PRI) of each pulse signal is taken from [4, 100]. The PRI of the conventional pulse is randomly selected and remains unchanged in the above interval; the PRI of the staggered pulse is randomly taken as 2 - 5, and the minimum value is not less than 80% of the maximum value; the PRI of the group - variable and agile pulse is randomly taken as 3 - 4, and the minimum value is not less than 80% of the maximum value, and each PRI is randomly repeated 15 - 30 times; the PRI of the sliding - variable pulse is also randomly selected in the above interval, and its step size increases by 1 each time, with a total increase of 34 times; the PRI of the jittering pulse is randomly selected and remains unchanged in the above interval, and the jitter rate of its pulse signal is 15%. At the same time, add less than 1% of jitter to the signals other than the jitter - type pulse signals. The data length of each pulse sequence is 3000 units, and each unit represents , 1000 of each signal are generated. According to the different degrees of missing and false, the interference degree is at intervals of 10%, and there are a total of 28 cases. Therefore, a total of 140000 data are generated in the training set, and the test set also generates 28 data sets according to different degrees of interference. 200 of each type of signal are generated in each data set, and 1000 are generated in each type of data set.

[0029] S2. Preprocess the generated radar emitter pulse sequence data. Perform a discrete Fourier transform on the radar pulse sequence with interference data to obtain a frequency-domain signal with noise. Set the points less than 10% of the maximum amplitude to 0 in the frequency spectrum to achieve the denoising effect and obtain the denoised frequency-domain signal. Perform an inverse Fourier transform on the denoised frequency-domain signal and use 0.5 as the amplitude reference to obtain the denoised binary time-domain signal.

[0030] S3. Construct a TCN network and set the TCN network parameters.

[0031] S4. Input the time-domain and frequency-domain signals of the denoised radar emitter pulse sequence, i.e., the denoised signals described in step S2, into the TCN network in step S3. After feature extraction, splice and fuse the feature vectors of the two, and input them into two fully connected layers for combination, abstraction, transformation, and non-linear mapping.

[0032] S5. Iterate the network in the "training - testing" manner. When the number of iterations reaches n times, the training ends, and the network output is the predicted class label of the radar emitter pulse interval debugging method.

[0033] In step S1, a database program can be established through Matlab to simulate and generate a data set.

[0034] In step S2, the radar emitter pulse signal preprocessing steps include:

[0035] S01. Perform a discrete Fourier transform on the radar pulse sequence with interference data to obtain a frequency-domain signal with noise. Set the points less than 10% of the maximum amplitude to 0 in the frequency spectrum to achieve the denoising effect and obtain the denoised frequency-domain signal.

[0036] S02. Perform an inverse Fourier transform on the denoised frequency-domain signal and use 0.5 as the amplitude reference to obtain the denoised binary time-domain signal.

[0037] In step S3, set the TCN network parameters: set the output channels of the dilated causal convolution to [16, 32, 32, 32, 16], set the number of network layers to 5, set dropout to 0.1, set the kernel size to 5, set the loss function of the TCN network to the negative log-likelihood loss function, use the AdamOptimizer algorithm to control the learning rate, use the rectified linear unit function as the activation function, and use the LogSoftmax classifier as the output layer of the network.

[0038] The neurons in the TCN network are represented as follows:

[0039] (1)

[0040] In the formula Representative filter; Representative sequence; d represents the dilation factor; t represents time; k represents the kernel size.

[0041] The residual block in the TCN network is expressed as follows:

[0042] (2)

[0043] In the formula Represents the activation function; Represents the input sequence; Represents the sequence after operations such as convolution on the input sequence.

[0044] The rectified linear unit activation function is expressed as follows:

[0045] (3)

[0046] In the formula, Represents the input value, which is a constant coefficient, Represents the output of the rectified linear unit activation function.

[0047] The LogSoftmax classifier is expressed as follows:

[0048] (4)

[0049] In the formula, Represents the i-th element in the data, Represents taking the exponential of the i-th element, Represents taking the sum of the exponentials of all elements, Represents the scoring value output for the i-th element.

[0050] In the step S1, the typical inter-pulse modulation methods of radar radiation sources include conventional modulation method, staggered modulation method, group variable and agile modulation method, sliding modulation method, and jitter modulation method.

[0051] According to the second aspect of the object of the present invention, the present invention is based on an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. It is characterized in that when the processor executes the program, it implements the steps of the above-mentioned radar pulse signal recognition method based on the TCN network.

[0052] The pulse repetition interval (PRI) of each pulse signal generated by simulation is taken from the range [4, 100]. The PRI of the conventional pulse is randomly selected from the above interval and remains unchanged; the PRI of the staggered pulse is randomly taken from 2 to 5, and the minimum value is not less than 80% of the maximum value; the PRI of the group-variable agile pulse is randomly taken from 3 to 4, and the minimum value is not less than 80% of the maximum value, and each PRI is randomly repeated 15 to 30 times; the PRI of the sliding pulse is also randomly selected from the above interval, and its step size increases by 1 each time, with a total increase of 34 times; the PRI of the jitter pulse is randomly selected from the above interval and remains unchanged, and the jitter rate of its pulse signal is 15%. At the same time, less than 1% of jitter is added to the remaining signals except the jitter-type pulse signals. The data length of each pulse sequence is 3000 units, and each unit represents , 1000 of each type of signal are generated. According to the different degrees of missing and false, the interference degree is at intervals of 10%, and there are a total of 28 cases. Therefore, the training set generates a total of 140,000 pieces of data. The test set also generates 28 data sets according to different degrees of interference. 200 of each type of signal are generated in each data set, and 1000 are generated in each type of data set.

[0053] Among many deep learning models, the recurrent neural network RNN (Recurrent Neural Network) introduces the concept of time series into the network structure design, making it perform well in time series data analysis. However, for long-term sparse sequences with missing and false interferences, the RNN network is extremely prone to problems such as gradient explosion / vanishing. The TCN network considered in this invention adopts a structure of dilated causal convolution and residual connection based on convolutional operations, which can enable the neural network to extract the overall features of long-term sequences, and at the same time denoise in the frequency domain and restore the modulation features of the pulse sequence, thus realizing the correct recognition of different modulation methods.

[0054] Example 1:

[0055] S1. Use Matlab to establish a database program, simulate and generate data. The data set includes the above 5 types of radar signals. Among them, 28 groups of each type of signal are generated from the interference degree of 0% to 90% of missing and false, and 5000 pieces of data in each group are used as the training set. Another 28 groups are generated according to the interference degree, and 1000 pieces in each group are used as the test set. Add labels according to different modulation methods and save them as data files and label files respectively.

[0056] S2. Perform discrete Fourier transform on the generated radar signals, set the points with amplitudes less than 10% of the maximum amplitude to 0 in the frequency spectrum to obtain the denoised frequency-domain signal, and perform inverse Fourier transform on it to obtain the denoised time-domain binary radar pulse sequence with 0.5 as the benchmark.

[0057] S3. Construct a TCN network. Set the TCN neurons:

[0058] Set the number of neurons in the TCN network to [16, 32, 32, 32, 16], where the neurons are represented as follows:

[0059]

[0060] Set the loss function of the TCN network: In this example, the negative log-likelihood loss function is selected, but it is not limited to the log-likelihood loss function.

[0061] Set the optimization algorithm of the TCN network: In this example, the Adam algorithm is selected, but it is not limited to the Adam algorithm.

[0062] Set the activation function of the TCN network: In this example, the rectified linear unit activation function is selected, but it is not limited to the rectified linear activation function. The rectified linear activation function is represented as follows:

[0063]

[0064] S4. Concatenate and fuse the time-domain and frequency-domain feature vectors of the signal after feature extraction, and input them into the fully connected layer for combination, abstraction, transformation, and non-linear mapping after dimensionality reduction.

[0065] S5. Set the number of iterations of the TCN network to 50, the learning rate to 0.0005, and the learning rate becomes 0.1 times the original after each iteration. Input the training samples and test samples after frequency-domain denoising into the set TCN network at the same time, and cycle and iterate the TCN network in the "training - testing" manner. When the number of iterations reaches 50, end the cycle iteration, and output the predicted category of each test data and the accuracy of successful signal recognition after each iteration.

[0066] Through the above-described embodiments, those skilled in the art can clearly understand that the facilities of the present invention can be implemented by means of software plus a necessary general hardware platform. The embodiments of the present invention can be implemented using existing processors, or by a dedicated processor used for this purpose or other purposes in a suitable system, or by a hardwired system. The embodiments of the present invention also include non-transitory computer-readable storage media, which include machine-readable media for carrying or having machine-executable instructions or data structures stored thereon; such machine-readable media can be any available medium accessible by a general or dedicated computer or other machine having a processor. For example, such machine-readable media can include RAM, ROM, EPROM, EEPROM, CD-ROM or other optical disk memories, magnetic disk memories or other magnetic storage devices, or any other medium that can be used to carry or store the required program code in the form of machine-executable instructions or data structures and can be accessed by a general or dedicated computer or other machine with a processor. When information is transmitted or provided to a machine through a network or other communication connection (hardwired, wireless or a combination of hardwired and wireless), the connection is also regarded as a machine-readable medium.

[0067] Emulation

[0068] The emulation generates five typical radar pulse signals, including conventional pulse signals, staggered pulse signals, group-variable agile pulse signals, sliding pulse signals, and jittered pulse signals. The pulse repetition interval (PRI) of each pulse signal is taken from [4, 100]. The PRI of the conventional pulse is randomly selected from the above interval and remains unchanged; the PRI of the staggered pulse is randomly taken as 2 - 5, and the minimum value is not less than 80% of the maximum value; the PRI of the group-variable agile pulse is randomly taken as 3 - 4, and the minimum value is not less than 80% of the maximum value, and each PRI is randomly repeated 15 - 30 times; the PRI of the sliding pulse is also randomly selected from the above interval, with a step size increase of 1 each time, and a total increase of 34 times; the PRI of the jittered pulse is randomly selected from the above interval and remains unchanged, and the jitter rate of its pulse signal is 15%.

[0069] To conform to the actual situation, less than 1% of jitter is added to the remaining signals except for the jitter-type pulse signals. The data length of each pulse sequence is 3000 units, and each unit represents , 1000 of each signal are generated. According to the different degrees of missing and false, the interference degree is at intervals of 10%, and there are a total of 28 cases. Therefore, a total of 140000 data are generated in the training set. The test set also generates 28 data sets according to different degrees of interference. In each data set, 200 of each type of signal are generated, and 1000 are generated in each type of data set.

[0070] II. Data preprocessing

[0071] Discretely Fourier transform the radar pulse sequence with interfering data to obtain the frequency-domain signal with noise. Set the points less than 10% of the maximum amplitude to 0 in the frequency spectrum to obtain the denoised frequency-domain signal. Perform inverse Fourier transform on the denoised frequency-domain signal, and use 0.5 as the amplitude reference to obtain the denoised binary time-domain signal.

[0072] III. Constructing the Network

[0073] Set the TCN network parameters: Set the output channels of the dilated causal convolution to [16, 32, 32, 32, 16], set the number of network layers to 5, set dropout to 0.1, set the kernel size to 5, set the loss function of the TCN network to the negative log-likelihood loss function, use the AdamOptimizer algorithm to control the learning rate, use the rectified linear unit function as the activation function, and use the LogSoftmax classifier as the output layer of the network.

[0074] IV. Data Reprocessing

[0075] By inputting the denoised time-domain and frequency-domain signals into the TCN network respectively, obtain their feature vectors, concatenate and fuse them and input them into two fully connected layers for combination, abstraction, transformation and non-linear mapping, and finally obtain the probability distribution of the pulse sequence. The parameters of the fully connected layers are (96000, 6000) and (6000, 5) respectively.

[0076] IV. Experimental Results

[0077] As described above, the experimental results are shown in the following table

[0078] Table 1 Recognition accuracy of the network under different working conditions

[0079]

[0080] Conclusion

[0081] The present invention proposes a method for identifying the pulse-to-pulse modulation mode of radar radiation sources based on a TCN network, which uses the convolution characteristics and the time-frequency characteristics of the sequence to solve the problem of different degrees of false and missing rates and the difficulty of identifying the pulse-to-pulse modulation mode of long-time sparse radar radiation sources when traditional algorithms fail. The simulation experiment verifies the effectiveness and robustness of this method. The simulation results show that the recognition accuracy of the network is 99.8% for the radar pulses of 5 different modulation methods without interfering data, and the recognition rates are 97.4% and 98.2% respectively under the interference degree of 90% of missing and false, and the TFT network can still maintain a high recognition rate under the extreme working condition of 90% interference degree.

Claims

1. A method for identifying the inter-pulse modulation mode of radar radiation sources based on a TCN network, characterized in that: The steps are as follows: S1. Preprocess the signals of the training set and the test set: S01. Perform a discrete Fourier transform on the radar pulse sequence with interference data to obtain a frequency-domain signal with noise. Set the points less than 10% of the maximum amplitude to 0 in the frequency spectrum to achieve the effect of noise reduction and obtain the denoised frequency-domain signal; S02. Perform an inverse Fourier transform on the denoised frequency-domain signal, and use 0.5 as the amplitude reference to obtain the denoised binary time-domain signal; S2. Construct a TCN network and set the TCN network parameters: S01. Set the number of output channels of the dilated causal convolution to [16, 32, 32, 32, 16], set the number of network layers to 5, set the dropout to 0.1, set the kernel size to 5, set the loss function of the TCN network to the negative log-likelihood loss function, use the AdamOptimizer algorithm to control the learning rate, use the rectified linear unit function as the activation function, and use the LogSoftmax classifier as the output layer of the network; S02. The neurons in the TCN network are represented as follows: where F = (f 1 , f 2 ,.., f K ) represents the filter; X = (x 1 , x 2 ,.., x T ) represents the sequence; d represents the dilation factor; t represents time; k represents the kernel size; S03. The residual blocks in the TCN network are represented as follows: Where Activation() represents the activation function; x represents the input sequence; represents the sequence after the input sequence undergoes a convolution operation; S04. The rectified linear unit activation function is represented as follows: f(y) = max(0, y) (3) In the formula, y represents the input value, which is a constant coefficient, and f(y) represents the output of the rectified linear unit activation function; S05. The LogSoftmax classifier is represented as follows: where x i represents the i-th element in the data, exp(x i ) represents taking the exponential of the i-th element, ∑ j exp(x j ) represents taking the sum of the exponentials of all elements, LogSoftmax(x i ) represents the scoring value output for the i-th element; S3. Input the time-domain and frequency-domain signals of the denoised radar radiation source pulse sequence, that is, the denoised signals described in step S1, into the TCN network in step S2. After feature extraction, splice and fuse the feature vectors of the two, and input them into two fully connected layers for combination, abstraction, transformation, and non-linear mapping; S4. Iterate the network in the "training - testing" manner. When the number of iterations reaches n times, the training ends, and the network output is the predicted category label of the radar radiation source inter-pulse debugging method.

Images