A known radar signal sorting method based on the gated recurrent network principle

Abstract

The present invention is a known radar signal sorting method based on the principle of gated recurrent networks. The scheme is as follows: signals are classified based on a known radar database to generate an original dataset; the original PDW stream data is processed to construct a joint feature vector; the processed joint feature vector is divided into a training sample set and a test sample set; a radar signal sorting network based on the principle of gated recurrent networks is constructed; the network is trained using the training sample set according to a set loss function and optimization method to obtain a trained sorting network; and the trained sorting network is tested using the test sample set. Using only PDW stream data, the present invention can effectively extract the timing characteristics of the pulses before and after the PDW stream, and can also obtain relatively accurate classification results in high pulse loss environments, with good pulse loss resistance.

Description

Technical Field

[0001] The present invention belongs to the field of signal processing, and further relates to a known radar signal sorting method based on the gated recurrent (GRU) network principle. Background Art

[0002] Radar signal sorting is a key step in signal processing in electronic reconnaissance systems. External radar signals are measured and processed in the receiver, and radar pulse information is transmitted to the back-end data processing module in the format of radar pulse descriptor words (PDWs). Using PDW stream information for radar signal sorting has obvious engineering application value and is a key research focus in electronic reconnaissance system engineering.

[0003] With the rapid development of various new radar systems, pulse radar signal parameters vary in complex and diverse patterns. The density of signal pulses entering reconnaissance receivers has increased dramatically, leading to interference and aliasing between signals, as well as a large number of false pulses and pulse loss. Due to the diverse and wide range of parameter variations in new radar systems, traditional template matching methods are unable to cope with these variable radar signal parameters. They can only passively address these agile and jagged parameter variations by increasing the tolerance. However, this increased tolerance allows many non-target radar signals to pass through, resulting in classification errors. Traditional template matching methods cannot meet the needs of pre-sorting complex, known radar signals. However, when known classification data is available, machine learning techniques can achieve excellent results. Neural networks can be used to perform supervised learning on radar database data to complete known signal sorting.

[0004] The changes in radar parameters clearly have temporal correlations, and recurrent neural networks, a broad category of neural networks, have unique analytical capabilities for time-series related data. To improve training speed and reduce computing overhead, and given the limited availability of known radar signal data samples, a gated recurrent unit network model is used to construct a sorting network. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a known radar signal sorting method based on the GRU network principle that can address the defects existing in the prior art. This solution can extract the timing characteristics of the previous and next pulses when using PDW data, and can resist pulse loss, thereby achieving better known radar signal sorting accuracy.

[0006] Technical solution: The present invention provides a known radar signal sorting method based on the gated recurrent network principle, comprising the following steps:

[0007] Step 1: Generate an original radar pulse descriptor word (PDW) stream dataset based on radar signal parameters in a known radar database, and generate interference signal and unknown radar signal datasets with reference to the database.

[0008] Step 2: Process the original PDW stream data and combine the three parameters of the PDW stream: carrier frequency, pulse width, and pulse repetition interval to establish a joint feature. To facilitate network learning, the joint feature group needs to be further normalized, rounded, one-hot encoded, and compressed and concatenated to obtain the joint feature input vector.

[0009] Step 3: Divide the obtained joint feature input vector set into a training data set, a validation data set, and a test data set according to the ratio of 75%, 5%, and 20% respectively;

[0010] Step 4: Construct a radar signal sorting network based on the gated recurrent network principle. The network consists of a gated recurrent GRU network and a fully connected network. The joint feature input vector is extracted in the gated recurrent GRU network to identify possible temporal and numerical change associations. The gated recurrent GRU learns these pattern change domain associations and stores them in the hidden layer. The classification task is completed by connecting the fully connected layer output by the gated recurrent GRU network.

[0011] Step 5: Use the training sample set to train the network according to the set loss function and optimization method to obtain a trained sorting network;

[0012] Step 6: Use the test sample set to test the trained sorting network, complete signal sorting, and calculate the accuracy.

[0013] in,

[0014] In step 1, the radar signal parameters in the known radar database are mainly referenced as follows: carrier frequency in MHz and modulation type, pulse width in μs, pulse repetition interval PRI in μs and modulation type; in addition, there is a radar serial number to identify different radars. The parameters of the generated interference and unknown signals are selected to be similar to the known radar signal parameters.

[0015] The modulation mode of the carrier frequency is selected to be fixed, jitter, group change or agile, and the modulation mode of the pulse repetition interval PRI includes fixed, jitter, staggered, sliding, group change and agile.

[0016] The step 2 specifically includes:

[0017] Step 2.1: Combine the three parameters of the radar signal PDW, namely, the carrier frequency RF, the pulse width parameter PW, and the pulse repetition interval, to create a joint feature vector. The joint feature vector is first expressed in a form similar to the PDW sequence, as shown in Equation (1):

[0018] {PRI1,PW1,RF1},...,{PRI m ,PW m ,RFm} (1)

[0019] The parameters need to be set within a limited range, namely the maximum pulse repetition interval value max_PRI and the maximum pulse width max_PW. When the maximum range is exceeded, the sequence is split into two different sequence streams;

[0020] Step 2.2: Perform digital rounding. First, realize dimensionless transformation and set the quantization processing range of each parameter. Each parameter is compared with the quantization processing range to remove the dimension unit. The numerical ranges with great differences between different parameter characteristics can be narrowed and unified. Then, the dimensionless characteristic parameters are digitally rounded, as shown in formula (2):

[0021]

[0022] Step 2.3: Perform one-hot encoding and adopt the embedding matrix mapping method. The embedding matrix mapping is E pdw , compressing the sparse unique code into a dense vector e pdw , its formula is as follows (3):

[0023] e pdwi =E pdwi ·pdwi (3)

[0024] Step 2.4: Concatenate the densely encoded feature vectors to obtain the joint feature input vector x and input it into the neural network. The specific concatenation form is as shown in formula (4):

[0025] x=[e pri ,e pw ,e rf ] (4).

[0026] In step 4, the gated recurrent GRU network in the sorting model is used as a hidden layer. Since the amount of radar sorting data is small, a single-layer gated recurrent GRU network is sufficient. The state vector ht of the trained gated recurrent GRU network is input into the fully connected network, and the output result is a probability vector for each known radar model classification, representing the probability that the pulse belongs to a certain radar model.

[0027] In step 5, the loss function adopts the cross entropy loss function, and the optimization of the signal sorting network loss function adopts the adaptive momentum estimation Adam algorithm.

[0028] In step 6, in order to evaluate the ability of the sorting network to deal with the pulse loss problem, data points are randomly extracted from the test set data in the range of 0 to 30% to obtain a test set under pulse loss, and then the trained known radar signal sorting network is used for sorting.

[0029] Beneficial effects: The present invention discloses a known radar signal sorting method based on the GRU network principle. Compared with the existing technology, it has the following beneficial effects: it extracts the timing features before and after the radar pulse by only using PDW data, and does not require additional complex processing at the front end of the receiver, which facilitates the transformation and upgrading of existing equipment; in the case where the parameters of the radar signal to be sorted overlap and the change pattern is complex, it effectively utilizes the extracted timing correlation features and old features, and can achieve better sorting accuracy in a high pulse loss environment. The computational overhead is lower than that of other neural networks, and the requirements for the quantity and quality of data are not high, which is conducive to engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Schematic diagram of the known radar signal sorting network based on GRU of the present invention.

[0031] Figure 2 The figure is a comparison of the precision rate of the present invention and the traditional algorithm under different pulse loss rates.

[0032] Figure 3 This is a comparison chart of the precision rate of the present invention with that of the GRU network and the LSTM network under different pulse loss rates. DETAILED DESCRIPTION

[0033] The technical solution of the present invention is further described below in conjunction with specific implementation methods and drawings.

[0034] This specific embodiment discloses a known radar signal sorting method based on the GRU network principle. The schematic diagram of the known radar signal sorting network based on GRU is shown in FIG. Figure 1 As shown:

[0035] The specific application steps are as follows:

[0036] Step 1: Generate the original radar PDW stream dataset based on the radar signal parameters in the known radar database, and generate the interference signal and unknown radar signal datasets with reference to the database;

[0037] Step 2: Process the original PDW stream data and combine the three parameters of the PDW stream: carrier frequency, pulse width, and pulse repetition interval to establish a joint feature. To facilitate network learning, the joint feature group needs to be further normalized, rounded, one-hot encoded, and compressed and concatenated to obtain the joint feature input vector.

[0038] Step 3: Divide the obtained joint feature input vector set into a training data set, a validation data set, and a test data set according to the ratio of 75%, 5%, and 20% respectively;

[0039] Step 4: Build a radar signal sorting network based on the gated recurrent network principle. The network consists of a GRU network and a fully connected network. The joint feature input vector is extracted in the GRU network to identify possible temporal and numerical change associations. The GRU learns these pattern change domain associations and stores them in the hidden layer. The classification task is completed by connecting to the fully connected layer of the GRU network output.

[0040] Step 5: Use the training sample set to train the network according to the set loss function and optimization method to obtain a trained sorting network;

[0041] Step 6: Use the test sample set to test the trained sorting network, complete signal sorting, and calculate the accuracy.

[0042] One-hot encoding, also known as single-bit encoding, uses N bits to encode N states. Each state has its own independent bit, and at any given time, only one of the bits is valid. That is, only one bit is 1, and the rest are zero. For example, male and female can be encoded as female 10 male 01.

[0043] In step 1, the radar signal parameters in the known radar database are mainly referenced as follows: carrier frequency in MHz and modulation type, pulse width in μs, pulse repetition interval PRI in μs and modulation type; in addition, there is a radar serial number to identify different radars. The parameters of the generated interference and unknown signals are selected to be similar to the known radar signal parameters.

[0044] The modulation mode of the carrier frequency is selected to be fixed, jitter, group change or agile, and the modulation mode of the pulse repetition interval PRI includes fixed, jitter, staggered, sliding, group change and agile.

[0045] The step 2 specifically includes:

[0046] Step 2.1: Combine the three parameters of the radar signal PDW, namely, the carrier frequency RF, the pulse width parameter PW, and the pulse repetition interval, to create a joint feature vector. The joint feature vector is first expressed in a form similar to the PDW sequence, as shown in Equation (1):

[0047] {PRI1,PW1,RF1},...,{PRI m ,PW m ,RF m} (1)

[0048] The parameters need to be set within a limited range, namely the maximum pulse repetition interval value max_PRI and the maximum pulse width max_PW. When the maximum range is exceeded, the sequence is split into two different sequence streams;

[0049] Step 2.2: Perform digital rounding. First, realize dimensionless transformation and set the quantization processing range of each parameter. Each parameter is compared with the quantization processing range to remove the dimension unit. The numerical ranges with great differences between different parameter characteristics can be narrowed and unified. Then, the dimensionless characteristic parameters are digitally rounded, as shown in formula (2):

[0050]

[0051] Step 2.3: Perform one-hot encoding and adopt the embedding matrix mapping method. The embedding matrix mapping is E pdw , compressing the sparse unique code into a dense vector e pdw , its formula is as follows (3):

[0052] e pdwi =E pdwi ·pdwi (3)

[0053] Step 2.4: Concatenate the densely encoded feature vectors to obtain the joint feature input vector x and input it into the neural network. The specific concatenation form is as shown in formula (4):

[0054] x=[e pri ,e pw ,e rf ] (4).

[0055] In step 4, the gated recurrent GRU network in the sorting model is used as a hidden layer. Since the amount of radar sorting data is small, a single-layer gated recurrent GRU network is sufficient. The state vector ht of the trained gated recurrent GRU network is input into the fully connected network, and the output result is a probability vector for each known radar model classification, representing the probability that the pulse belongs to a certain radar model.

[0056] In step 5, the loss function adopts the cross entropy loss function, and the optimization of the signal sorting network loss function adopts the adaptive momentum estimation Adam algorithm.

[0057] In step 6, in order to evaluate the ability of the sorting network to deal with the pulse loss problem, data points are randomly extracted from the test set data in the range of 0 to 30% to obtain a test set under pulse loss, and then the trained known radar signal sorting network is used for sorting.

[0058] In this embodiment, a total of 6 radars with different known parameters are set up. The parameters of the set radars overlap, and the characteristics adopt a variety of modulation methods. The specific parameter settings are shown in Table 1. Each known radar generates 2000 pulse sample data, constituting an original sample data set with a total sample number of 12,000.

[0059] Table 1 Known radar simulation parameter settings

[0060]

[0061] In addition to carrier frequency, pulse width, and PRI parameters, the dataset samples also include radar serial numbers as data labels. The original radar simulation signal dataset is divided into training, validation, and test datasets at a ratio of 75%, 5%, and 20%, respectively. Due to the small sample size and the limited number of hyperparameters required for the signal sorting network, a small validation set is sufficient for preliminary evaluation and network tuning. For the GRU network, the length of the densely encoded joint feature input vector is 48. The parameter settings for the GRU radar signal sorting network are shown in Table 2:

[0062] Table 2 Network parameters

[0063]

[0064] Set up the network according to the parameters in the table and learn the network using a training set of simulated radar data. Evaluate the effectiveness of the known radar signal sorting network using the validation set and adjust its parameters to obtain a trained sorting network. Test the sorting network using a simulated radar signal test set. To assess the sorting network's ability to handle pulse dropout, randomly extract data points from the test set at a range of 0-30% to obtain a test set with pulse dropout. The trained known radar signal sorting network is then used to sort the training set at different dropout rates. The confusion matrices of the sorting results for the training set with different dropout rates are shown.

[0065] The radar signal sorting method based on GRU is compared with the traditional template matching method. Using the same parameters, their accuracy is compared under different pulse loss rates. Figure 2 ; In addition, it is compared with the sorting network designed based on simple RNN network and LSTM network, and its precision is compared under different pulse loss rates, such as Figure 3 .

[0066] Based on the simulation results and analysis content, when faced with the situation where the parameters of the radar signals to be sorted overlap and have complex change patterns, the known radar signal sorting scheme proposed in this invention can effectively extract and analyze parameter associations, quickly and accurately complete the known radar signal pre-sorting task, and meet the needs of the known radar signal sorting algorithm to resist pulse loss.

[0067] The above content is only for explaining the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A known radar signal sorting method based on the gated recurrent network principle, characterized in that: The method comprises the following steps: Step 1: Generate an original radar pulse descriptor word (PDW) stream dataset based on radar signal parameters in a known radar database, and generate interference signal and unknown radar signal datasets with reference to the database. Step 2: Process the original PDW stream data and combine the three parameters of the PDW stream: carrier frequency, pulse width, and pulse repetition interval to establish a joint feature. To facilitate network learning, the joint feature group needs to be further normalized, rounded, one-hot encoded, and compressed and concatenated to obtain the joint feature input vector. Step 3: Divide the obtained joint feature input vector set into a training data set, a validation data set, and a test data set according to the ratio of 75%, 5%, and 20% respectively; Step 4: Construct a radar signal sorting network based on the gated recurrent network principle. The network consists of a gated recurrent GRU network and a fully connected network. The joint feature input vector is extracted in the gated recurrent GRU network to identify possible temporal and numerical change associations. The gated recurrent GRU learns these pattern change domain associations and stores them in the hidden layer. The classification task is completed by connecting the fully connected layer output by the gated recurrent GRU network. Step 5: Use the training sample set to train the network according to the set loss function and optimization method to obtain a trained sorting network; Step 6: Use the test sample set to test the trained sorting network, complete signal sorting, and calculate the accuracy.

2. The known radar signal sorting method based on the gated recurrent network principle according to claim 1 is characterized in that: In step 1, the radar signal parameters in the known radar database are mainly referenced as follows: carrier frequency in MHz and modulation type, pulse width in μs, pulse repetition interval PRI in μs and modulation type; in addition, there is a radar serial number to identify different radars. The parameters of the generated interference and unknown signals are selected to be similar to the known radar signal parameters.

3. The known radar signal sorting method based on the gated recurrent network principle according to claim 2 is characterized in that: The modulation mode of the carrier frequency is selected to be fixed, jitter, group change or agile, and the modulation mode of the pulse repetition interval PRI includes fixed, jitter, staggered, sliding, group change and agile.

4. The known radar signal sorting method based on the gated recurrent network principle according to claim 1 is characterized in that: The step 2 specifically includes: Step 2.1: Combine the three parameters of the radar signal PDW, namely, the carrier frequency RF, the pulse width parameter PW, and the pulse repetition interval, to create a joint feature vector. The joint feature vector is first expressed in a form similar to the PDW sequence, as shown in Equation (1): {PRI1,PW1,RF1},...,{PRI m ,PW m ,RF m } (1) The parameters need to be set within a limited range, namely the maximum pulse repetition interval value max_PRI and the maximum pulse width max_PW. When the maximum range is exceeded, the sequence is split into two different sequence streams; Step 2.2: Perform digital rounding. First, realize dimensionless transformation and set the quantization processing range of each parameter. Each parameter is compared with the quantization processing range to remove the dimension unit. The numerical ranges with great differences between different parameter characteristics can be narrowed and unified. Then, the dimensionless characteristic parameters are digitally rounded, as shown in formula (2): Step 2.3: Perform one-hot encoding and adopt the embedding matrix mapping method. The embedding matrix mapping is E pdw , compressing the sparse unique code into a dense vector e pdw , its formula is as follows (3): is pdwi =E pdwi ·month (3) Step 2.4: Concatenate the densely encoded feature vectors to obtain the joint feature input vector x and input it into the neural network. The specific concatenation form is as shown in formula (4): x=[e pri ,And pw ,And rf ] (4)。 5. The known radar signal sorting method based on the gated recurrent network principle according to claim 1 is characterized in that: In step 4, the gated recurrent GRU network in the sorting model is used as a hidden layer. Since the amount of radar sorting data is small, a single-layer gated recurrent GRU network is sufficient. The state vector ht of the trained gated recurrent GRU network is input into the fully connected network, and the output result is a probability vector for each known radar model classification, representing the probability that the pulse belongs to a certain radar model.

6. The known radar signal sorting method based on the gated recurrent network principle according to claim 1 is characterized in that: In step 5, the loss function adopts the cross entropy loss function, and the optimization of the signal sorting network loss function adopts the adaptive momentum estimation Adam algorithm.

7. The known radar signal sorting method based on the gated recurrent network principle according to claim 1 is characterized in that: In step 6, in order to evaluate the ability of the sorting network to deal with the pulse loss problem, data points are randomly extracted from the test set data in the range of 0 to 30% to obtain a test set under pulse loss, and then the trained known radar signal sorting network is used for sorting.

Images