Radar pri modulation type recognition method based on neural network and multi-head attention mechanism

Abstract

The application relates to a radar PRI modulation type recognition method based on a neural network and a multi-head attention mechanism and belongs to the field of radar signal processing. First-order difference is performed on a radar pulse TOA sequence to obtain sample data of a PRI sequence, a median filter is used for filtering processing on the sample data, the influence of missing or false pulses is reduced, a model fusing a multi-scale convolution squeeze-and-excitation neural network, a bidirectional gated recurrent unit neural network and a multi-head attention mechanism is constructed, the model can extract local fine features of data and capture long-term time sequence correlation features of data, the multi-head attention mechanism is used for integrating global information, more weights are applied to important time steps in the gated recurrent unit, and the model is trained and tested; results show that, compared with other neural network methods, the recognition rate of the six radar PRI modulation signals is higher under the condition that the missing pulse and the high proportion of stray pulses.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of radar signal processing, in particular to the classification and identification of radar signal inter-pulse modulation, and more particularly to a radar signal PRI modulation type identification method based on deep learning. BACKGROUND

[0002] In modern electronic warfare, the complexity and diversity of radar technology make the identification of radar signal pulse repetition interval (PRI) modulation patterns particularly important. Radar warning receiver (RWR) and electronic support (ESM) systems rely on the identification of PRI modulation types for emitter identification, which is crucial for model identification, type determination, operating mode identification, and threat assessment. By accurately identifying PRI modulation patterns, not only can an emitter identification database be constructed, but also important evidence can be provided for electromagnetic environment situation awareness and interference strategy development. With the development of new radar systems, the PRI modulation patterns of radar signals are becoming increasingly complex, including not only traditional fixed patterns but also complex modulation forms such as irregular, jitter, group variation, sliding, and sinusoidal modulation. These complex PRI modulation patterns pose a significant challenge to radar signal identification technology, requiring the continuous development of new identification methods and techniques. Effective PRI modulation pattern identification technology can improve the detection and identification capabilities of electronic warfare systems, thereby gaining the upper hand in modern battlefields.

[0003] Generally speaking, the identification methods of radar signal PRI modulation patterns can be divided into traditional methods based on conventional parameters and deep learning-based methods. Traditional identification methods mainly include histogram method, PRI transformation method, and their variants, which can achieve the classification and identification of simple system radars through pulse distribution quantity statistics or by combining other manually extracted features. Deep learning-based methods can more effectively process high-dimensional and complex structured data, automatically extracting rich feature representations. It has stronger generalization ability, can learn the potential distribution of data, and has scalability and flexibility, achieving better recognition results by splicing different types of neural network algorithms. Currently, scholars at home and abroad have proposed some radar PRI modulation mode identification methods based on deep learning, such as methods based on long short-term memory (LSTM) networks, convolutional neural network (CNN) methods, and time-domain convolutional neural network (TCN) methods. However, the recognition accuracy of these methods still needs to be improved in the case of high pulse loss rate and high false pulse proportion. SUMMARY

[0004] The present application provides a radar PRI modulation type identification method based on neural networks and multi-head attention mechanisms to solve the problems of relying on experience-based threshold selection and low recognition accuracy in existing radar PRI modulation type identification methods.

[0005] The technical solution adopted by the present invention comprises the following steps:

[0006] Step 1: Simulate and generate six types of inter-pulse modulation radar signals, including fixed modulation signal, staggered modulation signal, jitter modulation signal, group-variable modulation signal, sinusoidal modulation signal, and sliding modulation signal. Noise processing is performed on the above six types of inter-pulse modulation radar signals to add missing pulses and false pulses in different proportions to obtain the radar signal arrival time TOA sequence data set;

[0007] Step 2: Perform a differential calculation on the radar signal TOA sequence to obtain the radar signal's pulse repetition interval (PRI) sequence. Next, use a median filter to filter the noisy PRI sequence containing false and missing pulses. Then, all PRI sequences are divided into a training set and a test set.

[0008] Step 3: Design a multi-scale convolution squeeze excitation neural network MSCSE module to replace the general convolutional neural network module, and fuse the extracted local features with different scales and weights into a feature vector to input into the next layer of neural network;

[0009] Step 4: Based on step 3, a radar PRI modulation recognition algorithm model based on a neural network and a multi-head attention mechanism is constructed. This model includes an MSCSE module, a bidirectional gated unit BiGRU network module, and a multi-head attention mechanism MHA module. The BiGRU network simultaneously utilizes contextual features in both the forward and backward directions of the sequence. The MHA module highlights the key information of the BiGRU network at different time steps, enhancing the algorithm model's ability to focus on important information in the data.

[0010] In step 5, the generated PRI sequence training set is sent to the algorithm model for model training, the model's hyperparameters are adjusted, and the model's weights and biases are updated. Using the trained algorithm model, the test sets under different noise conditions are sent to the model for classification and recognition, and the recognition rate of the radar signal PRI modulation mode under different noise conditions is obtained.

[0011] In step 1 of the present invention, six PRI modulated radar signals are simulated and generated. The time interval between the arrival of adjacent pulses is the pulse repetition interval PRI, which is defined by the following function:

[0012] P={p1,p2,…,p N},p i =t i+1 -t i

[0013] Where P is the PRI sequence, {t i, i = 1, 2, ..., N + 1} is the arrival time of each pulse signal in the radar pulse train being analyzed, i is the data sequence number, and N is the number of radar pulse repetition intervals contained in the radar pulse train. Different radar systems have different PRI operating modes and modulation modes. In practical applications, due to the influence of noise, the PRI sequence has a high loss and spurious pulse rate. Random noise, including pulse missing noise and pulse spurious noise, is added to the TOA sequences of the generated six PRI modulated signals. The missing pulse sequence can be expressed as:

[0014]

[0015] Among them, {p′ n} indicates the received TOA sequence T n The new PRI sequence when j pulses are missing from the i-th pulse to the i+j-1 pulse, the false pulse sequence can be expressed as:

[0016]

[0017] {p′ n ′} represents a new PRI sequence with j pulses added after the i-th pulse, T i is the interval between the last false pulse and the next true pulse of the ith pulse.

[0018] The PRIs of different radar systems described in the present invention have various operating modes and modulation modes, including the following six PRI modulation types:

[0019] Fixed PRI modulation signals generally have a fixed value, and their fluctuation range does not exceed 1%;

[0020] In a staggered PRI modulated signal, multiple PRI values ​​appear in a periodic sequence;

[0021] In a jittered PRI modulated signal, the PRI sequence value is jittered around a constant value. The jitter range is usually randomly selected between 3% and 30%, usually following a Gaussian distribution.

[0022] The value of the slip modulation PRI sequence usually increases monotonically from an initial value until it reaches a maximum value, or decreases monotonically from the maximum value to the minimum value;

[0023] The numerical variation pattern of the sinusoidally modulated PRI sequence has a shape similar to that of a sine function and appears periodically;

[0024] Group-variable modulation is similar to staggered modulation, with several fixed values. The difference is that different PRI values ​​in group-variable modulation have a certain dwell time, and the dwell arrays of several PRIs also change periodically.

[0025] In step 2 of the present invention, {p′ n}, {p′ n The first-order difference calculation of each inter-pulse modulation TOA sequence of ′} is performed to obtain the PRI sequence, which is first normalized to avoid a large difference in the order of magnitude of different data and to speed up the calculation. The normalization process is expressed as:

[0026]

[0027] Among them, s max is the maximum value in the current sequence, s min is the minimum value in the current sequence. Then, each normalized PRI sequence is subjected to median filtering. The median filter can effectively suppress the noise caused by missing pulses and false pulses. For a window of size 2K+1, it is expressed as:

[0028] x i =median{y i-K ,y i-K+1 ,…,y i+K}

[0029] where x i It is the data point after filtering. The new PRI sequence data after median filter processing is divided into training set and test set. The ratio of training set to data set is 8:2.

[0030] The total number of samples in the dataset of the present invention is 60,000.

[0031] In step 3 of the present invention, a multi-scale convolution module MSCSE is designed that integrates the squeeze-excited attention mechanism. First, four convolution kernels of different sizes, 1×1, 1×3, 1×5, and 1×7, are used to process data. The model can capture features of different scales. Smaller convolution kernels focus on local fine-grained features, while larger convolution kernels can capture longer-range dependencies, avoiding information loss that may result from relying solely on a single-scale convolution kernel. The calculation formula for a convolution kernel size of 1×k can be expressed as:

[0032]

[0033] Among them, a represents the ath channel under the convolution kernel operation, Represents the convolution weight of the a-th channel of the convolution kernel, Indicates the corresponding offset, X t is a subsequence of the input data intercepted at time step t, Is the corresponding output result. After the convolution process, the output result Y is obtained. kAfter batch normalization, the output of each convolutional layer is concatenated to obtain the total output Y of a multi-scale convolution. The concatenation operation formula is as follows:

[0034] Y=Concat(Y1,Y3,Y5,Y7)

[0035] Subsequently, through the squeezing and excitation process, the module can adaptively adjust the weights of the feature channels, making the network pay more attention to important features. First, the squeezing process. The input feature map dimension is H*1*C, which represents the length, width and number of channels of the feature map respectively. Global average pooling is performed on each channel to compress the feature map of each channel into a value, that is, 1*1*C, to obtain a global description of each channel. The operation expression is as follows:

[0036]

[0037] Among them, F sq Represents the squeeze operation, Y(i) represents the input feature of the i-th channel, followed by the excitation operation, which inputs the channel description vector into a small fully connected network, consisting of two fully connected layers FC and an activation function Sigmoid, to generate the weight coefficient of each channel. The weight coefficient is the scale factor for adjusting the original feature map. The calculation process expression is as follows:

[0038] ω c =F ex (Z c ,W)=Sigmoid(W2(W1Z c +b1)+b2)

[0039] Among them, F ex Represents the excitation operation, W1 and b1 are the weight and bias of the first fully connected layer, W2 and b2 are the weight and bias of the second fully connected layer. Finally, the learned weight coefficient is weighted to each channel feature, and the result is added to the input feature to obtain the final output. The calculation expression is as follows:

[0040] Y'=ω c Y+Y

[0041] Afterwards, the result is processed with the P-ReLU activation function. P-ReLU is an improved activation function that introduces learnable parameters based on the ReLU activation function to avoid the inability to learn effective features in the negative value area, thereby enhancing the expressive power of the model. The calculation expression is as follows:

[0042]

[0043] X'=f(Y')

[0044] Among them, α is a learnable parameter, and then the result X' after the P-ReLU activation function is processed is sent to the next layer of the network.

[0045] In step 4 of the present invention, a radar PRI modulation recognition algorithm model is constructed. The preprocessed PRI sequence data first passes through two MSCSE layers to extract features at different time scales, and then passes through a pooling layer to downsample the features to reduce the data dimension while retaining important information, thereby further improving the computational efficiency. Then, after passing through two BiGRU layers, the model can better learn the temporal relationship in the sequence and capture the context information. Subsequently, after passing through the multi-head attention layer model, the model can focus on the most important part of the PRI sequence, making information transmission more efficient. The data is flattened through the Flatten layer. Finally, the extracted features are integrated through a fully connected layer, and the classification results are finally output through the Softmax layer, thereby achieving accurate recognition of different types of PRI modulation.

[0046] In step 5 of the present invention, the generated PRI sequence training set is sent to the algorithm model for model training. The training process includes setting hyperparameters and updating network weights. Assuming that the final prediction result of the model is p q , the loss function used is the root mean square error function, and the calculation expression is as follows:

[0047]

[0048] Where B is the total number of batch samples, which is the same as the value of the hyperparameter Batch_size, and l q is the label value of the qth sample, p q is the predicted value of the qth sample, and RMSE is the loss function value, which is used for backpropagation to update the model weights.

[0049] The model parameters of the present invention are set as follows: the optimizer uses Adam, the batch_size sample batch number is 128, the epochs iteration number is 50, and the Lr learning rate is 0.001. Afterwards, the trained prediction model is tested using the test set to verify that the model has high prediction accuracy.

[0050] The advantage of the present invention is that it provides a new algorithm model for identifying the PRI modulation mode of radar signals. The algorithm model mainly includes a multi-scale convolution squeeze excitation neural network MSCSE module, a bidirectional gated unit BiGRU network module and a multi-head attention mechanism MHA module. It is compared with the existing CNN-LSTM, TCN, and CNN network structures. Under the conditions of different missing pulses and false pulse ratios, different pulse-to-pulse modulation signals are tested to verify the effectiveness of the method. The method integrates the multi-scale convolution squeeze excitation neural network MSCSE, the bidirectional gated recurrent unit neural network and the multi-head attention mechanism, and can extract local refined features of the data and capture the long-term temporal correlation features of the data. By integrating global information through the multi-head attention mechanism, more weight can be applied to important time steps in the gated recurrent unit. The classification and recognition effect of the radar signal PRI modulation mode is better under the condition of a high proportion of missing pulses and false pulses, and accurate identification of the radar signal PRI modulation type in a noisy environment can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of the radar signal PRI modulation mode identification method in the present invention;

[0052] Figure 2 It is a schematic diagram of the PRI sequence of the fixed modulation signal modulation mode;

[0053] Figure 3 This is a schematic diagram of the PRI sequence of the jitter modulation signal modulation method;

[0054] Figure 4 This is a schematic diagram of the PRI sequence of the staggered modulation signal modulation method;

[0055] Figure 5 This is a schematic diagram of the PRI sequence of the group variable modulation signal modulation method;

[0056] Figure 6 This is a schematic diagram of the PRI sequence of the sinusoidal modulation signal modulation method;

[0057] Figure 7 This is a schematic diagram of the PRI sequence of the sliding modulation signal modulation method;

[0058] Figure 8 This is a schematic diagram of the proposed multi-scale convolutional neural network module structure with a fusion squeeze incentive mechanism;

[0059] Figure 9 This is the overall structural diagram of the proposed radar PRI modulation recognition algorithm model;

[0060] Figure 10 It is a schematic diagram of the gated recurrent unit structure;

[0061] Figure 11 It is a schematic diagram of the structure of a bidirectional gated recurrent unit;

[0062] Figure 12 This is a schematic diagram of the structure of the multi-head attention mechanism for hidden state features;

[0063] Figure 13 The recognition rates of six radar PRI modulation signals obtained by simulation are as follows:

[0064] Figure 14 This is a comparison chart of the recognition rates of CNN-LSTM, TCN, CNN and the method of the present invention under different noise conditions. DETAILED DESCRIPTION

[0065] The technical solution of the present invention is described in detail below in conjunction with specific embodiments. Figure 1 As shown, the present invention provides a radar signal PRI modulation type recognition method based on a neural network and a multi-head attention mechanism, and the specific steps are as follows:

[0066] Step 1: Simulate and generate six pulse-to-pulse modulation radar signals. The time interval between the arrival of adjacent pulses is the pulse repetition interval PRI, which is defined by the following function:

[0067] P={p1,p2,…,p N},p i =t i+1 -t i

[0068] Where P is the PRI sequence, {t i , i = 1, 2, ..., N + 1} is the arrival time of each pulse signal in the radar pulse train being analyzed, i is the data sequence number, and N is the number of radar pulse repetition intervals contained in the radar pulse train. Different radar systems may have different PRI operating modes and modulation modes. The following describes the six types of PRI modulation signals generated:

[0069] (1) Fixed PRI modulation type, see Figure 2 , the fixed PRI modulation signal has a fixed value, and its fluctuation range does not exceed 1%. Its signal law can be expressed as:

[0070] p i =a

[0071] Where a is a constant. The fixed PRI modulation type is generally used in conventional search and tracking radar tasks.

[0072] (2) Different PRI modulation types, see Figure 3In the staggered PRI modulation signal, multiple PRI values ​​appear in a periodic sequence, and the signal pattern can be expressed as:

[0073] p i =p u ,i=umodM

[0074] Where M refers to the number of PRIs in one cycle, u = 1, 2, ..., M. Each cycle is called a skeleton cycle. Staggered PRI signals are often used to eliminate blind speed in MTI radar (moving target detection radar);

[0075] (3) Jitter PRI modulation type, see Figure 4 The value of the jitter-modulated PRI sequence jitters around a constant value. The jitter range is usually randomly selected between 3% and 30%, and usually follows a Gaussian distribution. Its signal law can be expressed as:

[0076] p i =p0*(1+δ)

[0077] Where δ is the jitter amplitude percentage, and its value range is 3%≤|δ|≤30%. Jittering PRI modulated signals is generally used to counter interference and anti-reconnaissance;

[0078] (4) Sliding PRI modulation type, see Figure 5 The value of the slip-modulated PRI sequence increases monotonically from an initial value until it reaches its maximum value, or decreases monotonically from its maximum value to its minimum value. After reaching the maximum or minimum value, the PRI is immediately reset to its initial value, and the process repeats. The monotonically increasing signal law can be expressed as:

[0079]

[0080] Among them, v represents the rate of increase, l represents the length of a change cycle, and p min It is the minimum value within the cycle. The PRI variation range of sliding modulation is generally 10% to 30% above and below the center value. The purpose of sliding PRI modulation is generally to eliminate blind distance;

[0081] (5) Sinusoidal PRI modulation type, see Figure 6 , the numerical variation law of the sinusoidal modulation PRI sequence has a shape similar to that of a sine function and appears periodically. Its signal law can be expressed as:

[0082]

[0083] Where A is the amplitude of the modulation, θ is the frequency of the sine function, is the phase of the sine function, o is the number of data in a sine cycle, λ oIt is a value proportional to o and defines the sampling resolution. Sine PRI modulated signals are generally used for missile guidance to avoid obscuration and range finding.

[0084] (6) Group variable PRI modulation type, see Figure 7 ,Group-variable modulation is similar to staggered modulation, with several fixed values. The difference is that different PRI values ​​in group-variable modulation have a certain dwell time, and the dwell arrays of several PRIs also have periodic changes. Its signal law can be expressed as:

[0085] PRI(N)=PRI0+N*△PRI

[0086] Where PRI(N) represents the Nth group of PRI values, PRI0 represents the first group of PRI values, N represents the number of groups (counting starts from 0), and △PRI represents the increment of the pulse repetition interval between each group. Group-variable PRI modulation can analyze range and velocity ambiguity.

[0087] In practical applications, due to the influence of noise, the PRI sequence has a high rate of missing and spurious pulses. Random noise, including missing pulse noise and spurious pulse noise, is added to the TOA sequences of the six generated PRI modulated signals. The missing pulse sequence can be expressed as:

[0088]

[0089] Among them, {p′ n} indicates the received TOA sequence T n The new PRI sequence when j pulses are missing from the i-th pulse to the i+j-1 pulse, the false pulse sequence can be expressed as:

[0090]

[0091] {p′ n ′} represents a new PRI sequence with j pulses added after the i-th pulse, T i is the interval between the last false pulse and the next true pulse of the i-th pulse;

[0092] Step 2, include {p′ n}, {p′ n The first-order difference calculation of each inter-pulse modulation TOA sequence of ′} is performed to obtain the PRI sequence, which is first normalized to avoid a large difference in the order of magnitude of different data and to speed up the calculation. The normalization process is expressed as:

[0093]

[0094] Among them, s max is the maximum value in the current sequence, s minis the minimum value in the current sequence, s i is the i-th processed data point. Then, each normalized PRI sequence is subjected to median filtering. The median filter can effectively suppress the noise caused by missing pulses and false pulses. For a window of size 2K+1, it is expressed as:

[0095] x i =median{y i-K ,y i-K+1 ,…,y i+K}

[0096] where x i It is the data point after filtering. The new PRI sequence data after median filter processing is divided into training set and test set. The total number of samples in the data set is 60,000, and the ratio of training set to data set is 8:2.

[0097] Step 3: Design a multi-scale convolution module that integrates the squeeze-incentive attention mechanism. Its structure is as follows: Figure 8 As shown in the figure. First, four different sizes of convolution kernels, 1×1, 1×3, 1×5, and 1×7, are used to process the data. The model can capture features of different scales. Smaller convolution kernels focus on local fine-grained features, while larger convolution kernels can capture longer-range dependencies, avoiding the information loss that may result from relying solely on a single-scale convolution kernel. The calculation formula for a convolution kernel size of 1×k can be expressed as:

[0098]

[0099] Among them, a represents the ath channel under the convolution kernel operation, Represents the convolution weight of the a-th channel of the convolution kernel, Indicates the corresponding offset, X t is a subsequence of the input data intercepted at time step t, Is the corresponding output result. After the convolution process, the output result Y is obtained. k After batch normalization and maximum pooling, the output of each convolution layer is concatenated to obtain the total output Y of a multi-scale convolution. The concatenation operation formula is as follows:

[0100] Y=Concat(Y1,Y3,Y5,Y7)

[0101] Subsequently, through the squeezing and excitation process, the module can adaptively adjust the weights of the feature channels, making the network pay more attention to important features. First, the squeezing process. The input feature map dimension is H*1*C, which represents the length, width and number of channels of the feature map respectively. Global average pooling is performed on each channel to compress the feature map of each channel into a value, that is, 1*1*C, to obtain a global description of each channel. The operation expression is as follows:

[0102]

[0103] Among them, F sq Represents the squeeze operation, Y(i) represents the input feature of the i-th channel, followed by the excitation operation, which inputs the channel description vector into a small fully connected network, consisting of two fully connected layers FC and an activation function Sigmoid, to generate the weight coefficient of each channel. The weight coefficient is the scale factor for adjusting the original feature map. The calculation process expression is as follows:

[0104] ω c =F ex (Z c ,W)=Sigmoid(W2(W1Z c +b1)+b2)

[0105] Among them, F ex Represents the excitation operation, W1 and b1 are the weight and bias of the first fully connected layer, W2 and b2 are the weight and bias of the second fully connected layer. Finally, the learned weight coefficient is weighted to each channel feature, and the result is added to the input feature to obtain the final output. The calculation expression is as follows:

[0106] Y'=ω c Y+Y

[0107] Afterwards, the result is processed with the P-ReLU (Parametric ReLU) activation function. P-ReLU is an improved activation function that introduces learnable parameters based on the ReLU activation function to avoid the inability to learn effective features in the negative value area, thereby enhancing the expressive power of the model. The calculation expression is as follows:

[0108]

[0109] X'=f(Y')

[0110] Among them, α is a learnable parameter. Then, the result X' after the P-ReLU activation function is sent to the next layer of the network;

[0111] Step 4: Based on step 3, a radar PRI modulation recognition algorithm model based on neural network and multi-head attention mechanism is constructed. The overall structure is as follows: Figure 9 As shown in the figure, the preprocessed PRI sequence data first passes through two MSCSE layers to extract features at different time scales. It then passes through a pooling layer to downsample the features, reducing the data dimension while retaining important information and further improving computational efficiency. It then passes through two BiGRU layers, enabling the model to better learn the temporal relationships in the sequence and capture contextual information. The multi-head attention layer then allows the model to focus on the most important parts of the PRI sequence, making information transfer more efficient. Finally, a fully connected layer integrates the extracted features, and the Softmax layer outputs the classification results, thereby achieving accurate identification of different types of PRI modulation.

[0112] The main advantage of GRU is that it solves the long dependency problem in traditional RNN networks through the gating mechanism. Compared with LSTM, GRU structure is simpler and only contains update gate and reset gate. Figure 10 As shown. The update gate is responsible for controlling the information transfer between the hidden state at the current moment and the hidden state at the previous moment, which helps to capture long-term dependencies in the sequence. The reset gate controls how to combine the hidden state at the previous moment with the current input. When the reset gate is close to 0, GRU tends to ignore the previous state, allowing the network to pay more attention to the input information at the current moment, which helps to capture short-term dependencies. Assume that the input information of GRU at time step t is X t ', reset the gate to R t , update gate to Z t , the current hidden state is H t , the candidate hidden state is Update the calculation expression as follows:

[0113] R t =Sigmoid(W xr X t '+W hr H t-1 +b r )

[0114] Z t =Sigmoid(W xz X t '+W hz H t-1 +b z )

[0115]

[0116] Among them, W xr 、W hr 、W xz、W hz 、W xh and W hh is the weight update matrix, b r 、b z and b h is the bias, tanh is the activation function, and W is the Hadamard product operator.

[0117] BiGRU Figure 11 As shown in the figure, it consists of two GRUs with opposite directions, one responsible for forward propagation of information and the other responsible for backward propagation of information. It can simultaneously maintain the association between future information and past information, thus performing well in prediction and recognition. It can capture dependencies at different time scales and provide more contextual information.

[0118] The specific calculation expression is:

[0119]

[0120] in, is the hidden state at time step t in the positive direction, is the hidden state at time step t-1 in the positive direction, is the hidden state of the reverse direction t time steps, is the hidden state of the t-1 time step in the reverse direction, H t is the result of merging the hidden states of the two directions at t time steps. The hidden state output of each time step of BiGRU is merged and the result H is used as the input of the next layer of multi-head attention mechanism. The attention mechanism is used to further extract contextual information related to the target task and enhance the model's ability to focus on important information. Through multi-head parallel computing, the model can focus on different parts of the input sequence from multiple different angles, such as Figure 12 As shown, for the rth attention head, the query vector (Query) Q is first calculated r , key vector (Key) K r Sum value vector (Value) V r The linear transformation of in, is the three-vector linear transformation matrix for the r-th head, then the attention output of the i-th head is calculated as

[0121]

[0122] where d K is the key vector dimension. Then, the outputs of all attention heads are concatenated and a linear transformation is performed to obtain the final attention output. The calculation expression is:

[0123] Multihead(Q,K,V)=Concat(head1,head2,...,head h )W O

[0124] Where h is the number of attention heads, set to 8, and W O is the weight matrix of the multi-head output. Then, the output of the multi-head attention layer is flattened through the Flatten layer, and then the model's final prediction result, i.e., the radar PRI modulation type category label, is obtained through the fully connected layer and the Softmax function.

[0125] Step 5: Send the generated PRI sequence training set into the algorithm model for model training. The training process includes setting hyperparameters and updating network weights. Assume that the final prediction result of the model is p q , the loss function used in this invention is the root mean square error function, and the calculation expression is as follows:

[0126]

[0127] Where B is the total number of batch samples, which is the same as the value of the hyperparameter Batch_size, and l q is the label value of the qth sample, p q is the predicted value for the qth sample, and RMSE is the loss function value, which is used for backpropagation to update the model weights. The model parameters are set as follows: Adam optimizer, Batch_size sample batches of 128, Epochs iterations of 50, and Lr learning rate of 0.001. The trained prediction model is then tested on the test set to verify its high prediction accuracy.

[0128] The beneficial effects of the present invention are further described below in conjunction with simulation experiments and results. The simulation tools for all simulation experiments are Matlab and Python 3.8.

[0129] First, the modulation parameters and added noise for the six PRI sequence data sets are introduced: the PRI value range is 100 to 3000 μs, the change rate of sliding modulation is 20 to 100 μs, the number of different PRIs for staggered modulation ranges from 3 to 10, the number of PRI switches for group modulation ranges from 3 to 10, the jitter range for jitter modulation ranges from 3% to 30%, and the amplitude variation range for sinusoidal modulation ranges from 10% to 20%. The pulse loss ratio is set from 0 to 50% with a simulation step size of 10%, and the false pulse ratio is set from 0 to 40% with a simulation step size of 8%. A total of six non-ideal reception scenarios are set, ranging from 0 to 5.

[0130] Experiment 1: The recognition rates of six radar PRI modulation signals were compared in six scenarios using the method of the present invention. Figure 13 As can be seen, the proposed method achieves recognition accuracy exceeding 90% for all six radar PRI modulation modes when the pulse loss ratio is within 50% and the false pulse ratio is within 40%. It can also accurately identify signals when the pulse loss ratio is within 30% and the false pulse ratio is within 24%, validating the effectiveness of the proposed method.

[0131] Experiment 2: The algorithm of the present invention is compared with the methods of CNN+LSTM, TCN and CNN algorithms, and the performance is analyzed based on the simulation results. Figure 14 It can be seen that the recognition effect of the algorithm of the present invention is better than other classic neural network methods under high noise conditions. Compared with the CNN+LSTM method with better recognition effect, the training time of this method is 2944s and the single sample test time is 9.4ms; the training time of the method of the present invention is 2523s and the single sample test time is 8.5ms, which verifies the high efficiency of the designed new algorithm model.

Claims

1. A radar PRI modulation type recognition method based on neural network and multi-head attention mechanism, characterized in that: The following steps are involved: Step 1: Simulate and generate six types of inter-pulse modulation radar signals, including fixed modulation signal, staggered modulation signal, jitter modulation signal, group-variable modulation signal, sinusoidal modulation signal, and sliding modulation signal. Noise processing is performed on the above six types of inter-pulse modulation radar signals to add missing pulses and false pulses in different proportions to obtain the radar signal arrival time TOA sequence data set; Step 2: Perform a differential calculation on the radar signal TOA sequence to obtain the radar signal's pulse repetition interval (PRI) sequence. Next, use a median filter to filter the noisy PRI sequence containing false and missing pulses. Then, all PRI sequences are divided into a training set and a test set. Step 3: Design a multi-scale convolution squeeze excitation neural network MSCSE module to fuse the extracted local features with different scales and weights into a feature vector and input it into the next layer of neural network; Step 4: Build a radar PRI modulation recognition algorithm model based on a neural network and a multi-head attention mechanism. This model includes an MSCSE module, a bidirectional gated unit (BiGRU) network module, and a multi-head attention mechanism (MHA) module. The BiGRU network simultaneously utilizes contextual features in both the forward and backward directions of the sequence. The MHA module highlights key information at different time steps of the BiGRU network, enhancing the algorithm model's ability to focus on important information in the data. In step 5, the generated PRI sequence training set is sent to the algorithm model for model training, the model's hyperparameters are adjusted, and the model's weights and biases are updated. Using the trained algorithm model, the test sets under different noise conditions are sent to the model for classification and recognition, and the recognition rate of the radar signal PRI modulation mode under different noise conditions is obtained.

2. The radar PRI modulation type recognition method based on neural network and multi-head attention mechanism according to claim 1 is characterized in that: In step 1, six PRI modulated radar signals are simulated and generated. The time interval between the arrival of adjacent pulses is the pulse repetition interval PRI, which is defined by the following function: P={p1,p2,…,p N },p i =t i+1 -t i Where P is the PRI sequence, {t i , i = 1, 2, ..., N + 1} is the arrival time of each pulse signal in the radar pulse train being analyzed, i is the data sequence number, and N is the number of radar pulse repetition intervals contained in the radar pulse train. Different radar systems have different PRI operating modes and modulation modes. In practical applications, due to the influence of noise, the PRI sequence has a high loss and spurious pulse rate. Random noise, including pulse missing noise and pulse spurious noise, is added to the TOA sequences of the generated six PRI modulated signals. The missing pulse sequence can be expressed as: Among them, {p′ n } indicates the received TOA sequence T n The new PRI sequence when j pulses are missing from the i-th pulse to the i+j-1 pulse, the false pulse sequence can be expressed as: {p′ n ′} represents a new PRI sequence with j pulses added after the i-th pulse, T i is the interval between the last false pulse and the next true pulse of the ith pulse.

3. The radar PRI modulation type recognition method based on neural network and multi-head attention mechanism according to claim 2 is characterized in that: The PRIs of different radar systems have different operating modes and modulation modes, including the following six PRI modulation types: Fixed PRI modulation signals generally have a fixed value, and their fluctuation range does not exceed 1%; In a staggered PRI modulated signal, multiple PRI values ​​appear in a periodic sequence; In a jittered PRI modulated signal, the PRI sequence value is jittered around a constant value. The jitter range is usually randomly selected between 3% and 30%, usually following a Gaussian distribution. The value of the slip modulation PRI sequence usually increases monotonically from an initial value until it reaches a maximum value, or decreases monotonically from the maximum value to the minimum value; The numerical variation pattern of the sinusoidally modulated PRI sequence has a shape similar to that of a sine function and appears periodically; Group-variable modulation is similar to staggered modulation, with several fixed values. The difference is that different PRI values ​​in group-variable modulation have a certain dwell time, and the dwell arrays of several PRIs also change periodically.

4. The radar PRI modulation type recognition method based on neural network and multi-head attention mechanism according to claim 1 is characterized in that: In step 2, the n }, {p′ n The first-order difference calculation of each inter-pulse modulation TOA sequence of ′} is performed to obtain the PRI sequence, which is first normalized to avoid a large difference in the order of magnitude of different data and to speed up the calculation. The normalization process is expressed as: Among them, s max is the maximum value in the current sequence, s min is the minimum value in the current sequence. Then, each normalized PRI sequence is subjected to median filtering. The median filter can effectively suppress the noise caused by missing pulses and false pulses. For a window of size 2K+1, it is expressed as: x i =median{and i-K ,and i-K+1 ,…,and i+K } where x i It is the data point after filtering. The new PRI sequence data after median filter processing is divided into training set and test set. The ratio of training set to data set is 8:

2.

5. The radar PRI modulation type recognition method based on neural network and multi-head attention mechanism according to claim 4 is characterized in that: The total number of samples in the dataset is 60,000.

6. The radar PRI modulation type recognition method based on neural network and multi-head attention mechanism according to claim 1 is characterized in that: In step 3, a multi-scale convolution module MSCSE is designed to integrate the squeeze-excited attention mechanism. First, four convolution kernels of different sizes, 1×1, 1×3, 1×5, and 1×7, are used to process the data. The model can capture features of different scales. Smaller convolution kernels focus on local fine-grained features, while larger convolution kernels can capture longer-range dependencies, avoiding the information loss caused by relying solely on a single-scale convolution kernel. The calculation formula for a convolution kernel size of 1×k can be expressed as: Among them, a represents the ath channel under the convolution kernel operation, Represents the convolution weight of the a-th channel of the convolution kernel, Indicates the corresponding offset, X t is a subsequence of the input data intercepted at time step t, Is the corresponding output result. After the convolution process, the output result Y is obtained. k After batch normalization, the output of each convolutional layer is concatenated to obtain the total output Y of a multi-scale convolution. The concatenation operation formula is as follows: Y=Concat(Y1,Y3,Y5,Y7) Subsequently, through the squeezing and excitation process, the module can adaptively adjust the weights of the feature channels, making the network pay more attention to important features. First, the squeezing process. The input feature map dimension is H*1*C, which represents the length, width and number of channels of the feature map respectively. Global average pooling is performed on each channel to compress the feature map of each channel into a value, that is, 1*1*C, to obtain a global description of each channel. The operation expression is as follows: Among them, F sq Represents the squeeze operation, Y(i) represents the input feature of the i-th channel, followed by the excitation operation, which inputs the channel description vector into a small fully connected network, consisting of two fully connected layers FC and an activation function Sigmoid, to generate the weight coefficient of each channel. The weight coefficient is the scale factor for adjusting the original feature map. The calculation process expression is as follows: ω c =F ex (Z c ,W)=Sigmoid(W2(W1Z c +b1)+b2) Among them, F ex Represents the excitation operation, W1 and b1 are the weight and bias of the first fully connected layer, W2 and b2 are the weight and bias of the second fully connected layer. Finally, the learned weight coefficient is weighted to each channel feature, and the result is added to the input feature to obtain the final output. The calculation expression is as follows: Y'=ω c Y+Y Afterwards, the result is processed with the P-ReLU activation function. P-ReLU is an improved activation function that introduces learnable parameters based on the ReLU activation function to avoid the inability to learn effective features in the negative value area, thereby enhancing the expressive power of the model. The calculation expression is as follows: X'=f(Y') Among them, α is a learnable parameter, and then the result X' after the P-ReLU activation function is processed is sent to the next layer of the network.

7. The radar PRI modulation type recognition method based on neural network and multi-head attention mechanism according to claim 1 is characterized in that: In step 4, a radar PRI modulation recognition algorithm model is constructed. The preprocessed PRI sequence data first passes through two MSCSE layers to extract features at different time scales. Then, it passes through a pooling layer to downsample the features and reduce the data dimension while retaining important information, thereby further improving computational efficiency. Then, it passes through two BiGRU layers, so that the model can better learn the temporal relationship in the sequence and capture the context information. Subsequently, it passes through a multi-head attention layer model to focus on the most important part of the PRI sequence, making information transmission more efficient. The data is flattened through a Flatten layer. Finally, the extracted features are integrated through a fully connected layer, and the classification results are finally output through a Softmax layer, thereby achieving accurate recognition of different types of PRI modulation.

8. The radar PRI modulation type recognition method based on neural network and multi-head attention mechanism according to claim 1 is characterized in that: In step 5, the generated PRI sequence training set is sent to the algorithm model for model training. The training process includes setting hyperparameters and updating network weights. Assuming that the final prediction result of the model is p q , the loss function used is the root mean square error function, and the calculation expression is as follows: Where B is the total number of batch samples, which is the same as the value of the hyperparameter Batch_size, and l q is the label value of the qth sample, p q is the predicted value of the qth sample, and RMSE is the loss function value, which is used for backpropagation to update the model weights.

9. The radar PRI modulation type recognition method based on neural network and multi-head attention mechanism according to claim 8 is characterized in that: The model parameters are set as follows: Adam is used as the optimizer, Batch_size is 128, Epochs is 50, and Lr is 0.

001. Afterwards, the trained prediction model is tested on the test set to verify that the model has high prediction accuracy.

Images