Pulse time distribution based convolutional neural network emitter identification method

Abstract

The application provides a pulse time distribution-based convolutional neural network radiation source identification method, and belongs to the technical field of electronic countermeasures.The application performs pulse stream density estimation, period extension, sliding window and two-dimensional clustering processing on full-pulse radiation source signals with different pulse repetition interval characteristics, converts the full-pulse radiation source signals into an arrival time difference sequence diagram with fixed resolution, takes the arrival time difference sequence diagram as a pulse time distribution characteristic sample of the radiation source, and realizes identification by using a convolutional neural network.In network training, the arrival time difference sequence diagram is taken as a training sample to complete network training; after the training is completed, full-pulse data is mapped to the output of each convolutional pooling layer, so that the identification of radar radiation sources with different pulse repetition interval characteristics can be realized.

Description

Technical Field

[0001] This invention relates to the field of electronic countermeasures technology. Background Technology

[0002] Radar source identification technology is one of the key technologies in modern electronic warfare. Electronic reconnaissance equipment can collect full-pulse radiation source data during electronic intelligence reconnaissance. Based on the changing patterns and characteristics of parameters such as frequency, pulse width, and repetition interval of the radiation source descriptor in the full-pulse data, intelligence information such as the radiation source's operating mode and purpose can be analyzed and determined. Based on significant features, the platform and model of the radiation source can even be further identified. However, with the widespread application of active phased array technology, the increasing complexity of modern radar functions, and the fact that a single radar can inherit multiple functions, it faces challenges. It has multiple operating modes, and its frequency, pulse width, and repetition period parameters change rapidly and overlap significantly. Furthermore, the distribution range of various radar signal parameters varies greatly and is uneven, with frequencies ranging from 400MHz to 18GHz, pulse widths from 0.1 to 1000µs, and periods from 3µs to 10000µs, making parameter normalization difficult. This makes traditional radiation source parameter-based identification methods difficult to adjust parameter thresholds, leading to a sharp increase in false alarm and false alarm rates, becoming a major technical challenge in radiation source identification. Summary of the Invention

[0003] To improve the intelligent identification capability of radar radiation sources in complex electromagnetic environments, this invention proposes a radiation source identification method based on pulse time distribution using convolutional neural networks. By applying a convolutional neural network (CNN), the pulse time distribution variation characteristics of the radiation source signal are extracted to achieve radiation source identification.

[0004] This invention performs pulse current density estimation, periodic extension, sliding windowing, and two-dimensional clustering on full-pulse radiation source signals with different pulse repetition interval characteristics, transforming them into time-of-arrival (TOA) sequence maps with fixed resolution. These TOA sequence maps are then used as radiation source features, and a convolutional neural network is employed for identification. During network training, the TOA sequence maps are used as training samples. After training, the full-pulse data is mapped to the outputs of each convolutional pooling layer, enabling the identification of radar radiation sources with different pulse repetition interval characteristics. The technical solution for implementing this invention includes:

[0005] Step 1: Cluster the intercepted full-pulse data of radar radiation sources using azimuth, pulse width, and frequency parameters;

[0006] Step 2: Calculate the pulse current density ΔT on the clustered full pulse data and determine the pulse repetition interval level k;

[0007] Step 3: Set the arrival time difference upper limit PRI_k to the preset pulse repetition interval level k. max Lower limit of arrival time difference PRI_k min Observation time window length W k And the number of pixels N on the time axis of the time difference sequence image samples. x Number of pixels N on the arrival time difference axis y As a preset parameter, the time axis clustering interval ΔT is calculated. x The time difference of arrival axis clustering interval ΔT y The time-axis sliding window step interval Δt is used to map full-pulse data into time difference sequence images;

[0008] Step 4: Based on the duration T of the full pulse data, extend or truncate the full pulse data to twice the length of the observation time window;

[0009] Step 5: Perform observation window sliding cycles on the full-pulse data with time axis sliding window step interval Δt, and generate a set of arrival time difference sequence diagrams of the full-pulse data by two-dimensional clustering;

[0010] Step 6: Generate a set of uncorrelated random noise images and overlay them with the arrival time difference sequence image to improve the robustness of the neural network.

[0011] Step 7: Network training phase. Process the full pulse data according to steps 1 to 6 to generate a training sample set, which is used to train the full pulse recognition neural network.

[0012] Step 8: Network testing phase. Process the full pulse data according to steps 1 to 6 to generate a test sample set, and test the full pulse recognition neural network trained in step 5.

[0013] Step 9: Repeat steps 7 and 8 until a network model with satisfactory recognition performance and generalization ability is obtained;

[0014] Compared with the prior art, the present invention has the following significant advantages: (1) It utilizes deep neural networks to achieve automatic machine extraction of radiation source features, which increases the efficiency of extracting radiation source features and is beneficial to the subsequent iterative updates of the identification system; (2) The arrival time difference sequence diagram generated by two-dimensional clustering can fully reflect the graphical features of the pulse time distribution change law of the radiation source, while reducing the matrix dimension, significantly reducing the amount of computation of convolution and pooling processing in the radiation source identification process, and improving the learning efficiency of the neural network; (3) The convolutional neural network based on pulse time distribution is suitable for the identification of radar radiation sources with different pulse repetition interval features.

[0015] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0016] Figure 1 A block diagram of a convolutional neural network for radiation source identification based on pulse time distribution;

[0017] Figure 2 This is a sequence diagram of the arrival time difference after two-dimensional clustering of signal 1;

[0018] Figure 3 This is a sequence diagram of the arrival time difference after two-dimensional clustering of signal 2;

[0019] Figure 4 This is a sequence diagram of the arrival time difference after two-dimensional clustering of signal 3;

[0020] Figure 5 This is a sequence diagram of the arrival time difference after two-dimensional clustering of the signal 4.

[0021] Figure 6 This is a sequence diagram of the arrival time difference after two-dimensional clustering of signal 5;

[0022] Figure 7 This is a sequence diagram of the arrival time difference after two-dimensional clustering of the signal 6.

[0023] Figure 8 This is a sequence diagram of the arrival time difference after two-dimensional clustering of signal 7;

[0024] Figure 9 This is a sequence diagram of the arrival time difference after two-dimensional clustering of the signal 8.

[0025] Figure 10 This is a sequence diagram of the arrival time difference after two-dimensional clustering of signal 9;

[0026] Figure 11 This is a sequence diagram of the arrival time difference after two-dimensional clustering of signal 10;

[0027] Figure 12 This is a sequence diagram of the arrival time difference after two-dimensional clustering of signal 11;

[0028] Figure 13 This is a sequence diagram of the arrival time difference after two-dimensional clustering of signal 12;

[0029] Figure 14 A graph showing the number of iterations versus training / test accuracy.

[0030] Figure 15 This is a graph showing the number of iterations versus the loss function value. Detailed Implementation

[0031] The present invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments, but the scope of protection of the present invention is not limited by the implementing regulations.

[0032] Combination Figure 1This section describes a radiation source identification method based on convolutional neural networks using pulse time distribution. For example... Figure 1 As shown, the preferred implementation process of the present invention includes the following steps:

[0033] Step 1: Cluster the intercepted full-pulse data of radar radiation sources using azimuth, pulse width, and frequency parameters;

[0034] Step 2: Calculate the pulse current density ΔT on the clustered full pulse data and determine the pulse repetition interval level k;

[0035] Step 3: Set the arrival time difference upper limit PRI_k to the preset pulse repetition interval level k. max Lower limit of arrival time difference PRI_k min Observation time window length W k And the number of pixels N on the time axis of the time difference sequence image samples. x Number of pixels N on the arrival time difference axis y As a preset parameter, the time axis clustering interval ΔT is calculated. x The time difference of arrival axis clustering interval ΔT y The time-axis sliding window step interval Δt is used to map full-pulse data into time difference sequence images;

[0036] Step 4: Based on the duration T of the full pulse data, extend or truncate the full pulse data to twice the length of the observation time window;

[0037] Step 5: Perform observation window sliding cycles on the full-pulse data with time axis sliding window step interval Δt, and generate a set of arrival time difference sequence diagrams of the full-pulse data by two-dimensional clustering;

[0038] Step 6: Generate a set of uncorrelated random noise images and overlay them with the arrival time difference sequence image to improve the robustness of the neural network;

[0039] Step 7: Network training phase. Process the full pulse data according to steps 1 to 6 to generate a training sample set, which is used to train the full pulse recognition neural network.

[0040] Step 8: Network testing phase. Process the full pulse data according to steps 1 to 6 to generate a test sample set, and test the full pulse recognition neural network trained in step 5.

[0041] Step 9: Repeat steps 7 and 8 until a network model with satisfactory recognition performance and generalization ability is obtained;

[0042] Preferably, step 2 includes the following steps:

[0043] Step 2-1: For the n full-pulse data after clustering in Step 1, the pulse arrival times are t0, t1, t2...t n-1 The duration T of the full pulse data is T = t n-1 -t0, calculate the pulse current density as follows:

[0044] ΔT=n / T (1);

[0045] Step 2-2: Preset the number of pulse repetition interval levels to N, where the upper limit of the pulse repetition interval of the k-th level is PRI_k. max The lower limit of the pulse repetition interval is PRI_k min If the pulse current density ΔT described in step 2-1 satisfies the following condition:

[0046] PRI_k min ≤ΔT≤PRI_k max (2);

[0047] Then the pulse repetition interval corresponding to the full pulse data is considered to be the kth level.

[0048] Preferably, step 3 includes the following steps:

[0049] Step 3-1: Preset the observation window length of the k-th level pulse repetition interval to W. k The number of pixels on the time axis of the time difference sequence image is N. x The time axis clustering interval is calculated as follows:

[0050] ΔT x =W k / N x (3);

[0051] Step 3-2: Preset the number of pixels on the arrival time difference axis of the time difference sequence images to N. y Given the parameter PRI_k max PRI_k min The clustering interval of the arrival time difference axis is calculated as follows:

[0052] ΔT y =(PRI_k max -PRI_k min ) / N y (4);

[0053] Step 3-3: Preset a set of full-pulse data sliding window steps N s A secondary generation of time-difference sequence image sample set, based on known parameters W. k Calculate the time-axis sliding window step interval of the k-th level pulse repetition interval.

[0054]

[0055] In the formula, Δt is in μs. It is a floor operation.

[0056] Preferably, step 4 includes the following steps:

[0057] Step 4-1: Calculate the duration T of the full pulse data. If the condition is met...

[0058] T<2*W k (6);

[0059] The full-pulse data is then periodically extended, and the arrival times of the periodically extended full-pulse data are represented as t0, t1, t2...t m-1 The duration after extension is T. E =2*W k =t m-1 -t0;

[0060] Step 4-2: Calculate the duration T of the full pulse data. If the condition is met...

[0061] T>2*W k (7);

[0062] Otherwise, the full pulse data is truncated, and the arrival times of the truncated full pulse data are represented as t0, t1, t2...t m-1 The duration after truncation is T. E =2*W k =t m-1 -t0;

[0063] Preferably, step 5 includes the following steps:

[0064] Step 5-1: Perform sliding window processing on the full-pulse data after Step 4, and truncate an observation window of length W. k Given the data, the time axis sliding window step interval Δt, and the sliding window step N. s The second generation of time difference sequence image sample set, during the i-th sliding window, traverses the arrival time sequence t0, t1, t2...t of the full pulse data. m-1 If the arrival time t of the j-th pulse j Conditions met:

[0065] (Δt*i+t0)≤t j ≤(Δt*i+W k +t0) (8); then cache the full pulse data that meets the conditions to obtain the full pulse dataset of the i-th sliding window processing.

[0066] T i ={t|(Δt*i+t0)≤t≤(Δt*i+W k +t0)} (9);

[0067] Where i∈N, 0≤i≤N s ;

[0068] Step 5-2: Traverse the full pulse dataset after sliding window processing, and obtain N by two-dimensional clustering. s A two-dimensional matrix P(x) represents the time-arrival time difference. i ,y i For the full-pulse dataset T of the i-th sliding window processing... i Solve for the corresponding two-dimensional time-arrival time difference matrix P. i (x i ,y i ), where x i For time axis index, y i Index for arrival time difference axis:

[0069] x j =(t j -(Δt*i+t0)) / ΔT x (10);

[0070] y j =(t j -t j-1 ) / ΔT y (11);

[0071] Then the two-dimensional matrix P of time-arrival time difference i (x i ,y i ) Corresponding position (x j ,y j The element on the ) is incremented by 1.

[0072] P i (x j ,y j ) = P i (x j ,y j )+1 (12);

[0073] Preferably, step 6 includes the following steps:

[0074] Step 6-1: Randomly generate N s A mutually uncorrelated two-dimensional matrix P whose dimension is related to the time-arrival time difference. i (x i ,y i The same random noise value matrix N i (x i ,y i );

[0075] Step 6-2: Calculate the two-dimensional matrix of time-arrival time difference after adding noise.

[0076] P Noise_i (x j ,y j ) = P i (x j ,y j )+N i (x j ,y j (13);

[0077] Preferably, step 9 includes the following steps:

[0078] Step 9-1: Train the network using the training sample set as in Step 7 to obtain the network model N during the training process. train Computational network model N train Training accuracy δ on the training sample set train If the conditions are met:

[0079]

[0080] Then stop training and store the network model that has been trained. In the formula, δ represents the desired training accuracy that the network can achieve. εtrain For the network model N during the training process train The decision threshold for whether the training accuracy requirements are met;

[0081] Step 9-2: Calculate the network model The test accuracy δ of the test sample set in step 8 test ;

[0082] Step 9-3: Repeat steps 9-1 to 9-2. If the generalization condition is met:

[0083] |δ train -δ test |<δ ε (15); then stop network training and testing, and store the network model that has been trained at this time. In the formula, δ ε For the network model during training The threshold for determining whether the generalization requirement is met.

[0084] This invention provides a radiation source identification method based on pulse time distribution using a convolutional neural network. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

[0085] Example:

[0086] The analysis was conducted using training and testing sample sets consisting of radar full-pulse radiation source data with 12 different pulse repetition interval characteristics. The training sample set contained 9,600 time difference of arrival (TDOA) sequence maps, with 800 for each type of radiation source; the testing sample set contained 2,400 TDOA sequence maps, with 200 for each type of radiation source; and the validation sample set contained 1,200 TDOA sequence maps, with 100 for each type of radiation source. The CNN consists of two convolutional pooling layers, three fully connected layers, and a softmax classifier. The first layer has a 5×5 kernel size, one channel, six kernels, a stride of 1, and 2:1 max pooling. The second layer also has a 5×5 kernel size, six channels, six kernels, and 16 kernels, with a stride of 1 and 2:1 max pooling. The first fully connected layer has an output space dimension of 120, the second fully connected layer has an output space dimension of 80, and the third fully connected layer has an output space dimension of 12. The loss function is cross-entropy error, the optimization algorithm is Adam, the learning rate is 0.00025, the minibatch size is 64, and the delta function is... εtrain Set to 0.985, δ ε Set it to 0.025.

[0087] Figures 2 to 13 The time difference sequence diagrams obtained after two-dimensional clustering of the full pulse data of 12 types of signals are shown in the figure. After processing in steps 1 to 5, radiation source signals with different pulse repetition interval characteristics can generate time difference sequence diagrams with the same resolution. Figure 14 The curve represents the number of iterations versus training / test accuracy during network training. As the number of iterations increases, the training accuracy gradually increases and tends towards 1, eventually converging stably at δ. εtrain Numerically, the model achieves an average test accuracy of 99.12%, and after network convergence, the error between training accuracy and test accuracy does not exceed δ. ε . Figure 15 This is a curve showing the number of iterations versus the loss function value during network training. As the number of iterations increases, the loss function value gradually decreases, tending towards 0 until convergence. Figure 14 , Figure 15It can be seen that the network has the required recognition and generalization capabilities. The performance of the trained network can be verified using a validation sample set, and the recognition rate can reach over 98%.

Claims

1. A radiation source identification method based on convolutional neural networks with pulse time distribution, characterized in that: Step 1: Cluster the intercepted full-pulse data of radar radiation sources using azimuth, pulse width, and frequency parameters; Step 2: Calculate the pulse current density Δ on the clustered full pulse data. T Determine the pulse repetition interval level k ; Step 3: By pulse repetition interval level k Preset upper limit of arrival time difference PRI_k max Lower limit of arrival time difference PRI_ k min Observation time window length W k And the number of pixels on the time axis of the arrival time difference sequence map sample. N x Number of pixels on the arrival time difference axis N y As a preset parameter, the time axis clustering interval Δ is calculated. T x , arrival time difference axis clustering interval Δ T y Time axis sliding window step interval Δ t It is used to map full-pulse data into a time-of-arrival sequence graph; Step 4: Based on the duration of the full pulse data T Extend or truncate the full pulse data to twice the length of the observation time window; Step 5: Step the full-pulse data with a time-axis sliding window at intervals Δ t By performing a sliding cycle of the observation window, two-dimensional clustering is used to generate a set of arrival time difference sequence diagrams of full pulse data; Step 6: Generate a set of uncorrelated random noise images and overlay them with the arrival time difference sequence image to improve the robustness of the neural network; Step 7: Network training phase. Process the full pulse data according to steps 1 to 6 to generate a training sample set, which is used to train the full pulse recognition neural network. Step 8: Network testing phase. Process the full pulse data according to steps 1 to 6 to generate a test sample set, and test the full pulse recognition neural network trained in step 5. Step 9: Repeat steps 7 and 8 until a network model with satisfactory recognition performance and generalization ability is obtained.

2. The radiation source identification method based on pulse time distribution using convolutional neural networks according to claim 1, characterized in that, Step 2 also includes: Step 2-1: Clustering the data from Step 1 n The full pulse data, with pulse arrival times of [number] pulses respectively. t 0、 t 1. t 2…… t n-1 Full pulse data duration T for T=t n-1 - t 0, calculate the pulse current density as: D T=n / T (1); Step 2-2: The preset number of pulse repetition interval levels is N, where the first... k The upper limit of the pulse repetition interval is PRI_k max The lower limit of the pulse repetition interval is PRI_k min The pulse current density Δ in step 2-1 T If the conditions are met: PRI_k min ≤ D T ≤ PRI_k max (2); Then it is assumed that the pulse repetition interval corresponding to the full pulse data is the first pulse. k level.

3. The radiation source identification method based on convolutional neural network with pulse time distribution according to claim 1, characterized in that, Step 3 also includes: Step 3-1: Preset the first k The observation window length of the pulse repetition interval is W k The number of pixels on the time axis of the arrival time difference sequence map is N x The time axis clustering interval is calculated as follows: D T x =W k / N x (3); Step 3-2: Preset the number of pixels on the arrival time difference axis of the arrival time difference sequence map to be [number]. N y From known parameters PRI_k max , PRI_k min The clustering interval of the arrival time difference axis is calculated as follows: D T y = ( PRI_k max -PRI_k min ) / N y (4); Step 3-3: Preset a set of full-pulse data sliding window steps N s A second set of arrival time difference sequence graph samples is generated, based on known parameters. W k Calculate the first k Time-axis sliding window step interval for pulse repetition interval: （5）； In the formula, Δ t The unit is μs , It is a floor operation.

4. The radiation source identification method based on convolutional neural network with pulse time distribution according to claim 1, characterized in that, Step 4 also includes: Step 4-1: Calculate the duration of the full pulse data T If the conditions are met: T< 2 *W k （6）； Then, the full-pulse data is periodically extended, and the arrival time of the periodically extended full-pulse data is expressed as: t 0、 t 1. t 2…… t m-1 The duration after extension is T E = 2 *W k =t m-1 - t 0; Step 4-2: Calculate the duration of the full pulse data T If the conditions are met: T > 2 *W k （7）； The full pulse data is then truncated, and the arrival time of the truncated full pulse data is expressed as follows: t 0、 t 1. t 2…… t m-1 The duration after truncation is T E = 2 *W k =t m-1 - t 0.

5. The radiation source identification method based on convolutional neural network with pulse time distribution according to claim 1, characterized in that, Step 5 further includes: Step 5-1: Perform sliding window processing on the full-pulse data after Step 4, and truncate an observation window of length. W k The data, with the time axis sliding window step interval Δ known. t sliding window step N s The second generation of the arrival time difference sequence map sample set, the first i During the second sliding window, the arrival time series of the full pulse data is traversed. t 0、 t 1. t 2…… t m-1 If the first j Arrival time of each pulse t j Conditions met: (D t*i+t 0) ≤ t j ≤ (D t*i+W k +t 0) (8); Then, cache the full pulse data that meets the conditions to obtain the first... i Full-pulse dataset processed by secondary sliding window: （9）； in, ; Step 5-2: Traverse the full pulse dataset after sliding window processing, and obtain the results through two-dimensional clustering. N s A two-dimensional matrix of time-arrival time differences P ( x i , y i ), for the first i Full-pulse dataset processed by secondary sliding window T i Solve for the corresponding two-dimensional time-arrival time difference matrix. P i ( x i , y i ),in, x i For timeline index, y i Index for arrival time difference axis: （10）； （11）； Then the two-dimensional matrix of time-arrival time difference P i ( x i , y i Corresponding position ( x j , y j The element on the ) is incremented by 1: （12）。 6. The radiation source identification method based on pulse time distribution using convolutional neural networks according to claim 1, characterized in that: Step 6 also includes: Step 6-1: Randomly generate N s A set of independent two-dimensional matrices whose dimensions are related to the time-arrival time difference. P i ( x i , y i The same random noise value matrix N i ( x i , y i ); Step 6-2: Calculate the two-dimensional matrix of time-arrival time difference after adding noise: （13）。 7. The radiation source identification method based on pulse time distribution using convolutional neural networks according to claim 1, characterized in that: Step 9 further includes: Step 9-1: Train the network using the training sample set as in Step 7 to obtain the network model during the training process. N train Computational network model N train Training accuracy on the training sample set δ train If the conditions are met: （14）； Then stop training and store the network model that has been trained. In the formula, To achieve the desired training accuracy of the network, δ εtrain For the network model during training N train The decision threshold for whether the training accuracy requirements are met; Step 9-2: Calculate the network model Test accuracy of the test sample set in step 8 δ test ; Step 9-3: Repeat steps 9-1 to 9-2. If the generalization condition is met: （15）； Then stop network training and testing, and store the network model that has been trained at this point. In the formula, δ s For the network model during training The threshold for determining whether the generalization requirement is met.

Images