A joint signal detection and modulation identification method for low probability of intercept radars

Abstract

The application discloses a kind of low probability of interception radar joint signal detection and modulation identification processing method, can be received to the low probability of interception radar signal sequence containing multiple pulses, complex modulation type, flexible modulation parameter value, realize the accurate detection of existing pulse, and give the detection of each pulse corresponding start time and modulation type identification result;Based on the idea of multi-task learning, combined with long short-term memory network and cross attention mechanism, effectively extract input pulse hierarchical multi-scale time sequence features;Through the design of data preprocessing method based on sliding window segmentation, local hierarchical multi-scale feature extraction module for different time scales and global similarity mining module relying on cross attention mechanism;Similarity information between multiple pulses can be mined, the accumulation of multiple pulses is used to assist in completing the accurate detection of pulse and modulation type identification task, and effective signal detection and modulation identification are realized under low signal-to-noise ratio and non-ideal conditions.

Description

Technical Field

[0001] This invention relates to the field of radar electronic reconnaissance technology, specifically to a joint signal detection and modulation identification processing method for low probability of intercept radar. Background Technology

[0002] Low probability of intercept (LPI) radars typically possess characteristics such as wide frequency band, low power, complex modulation, and agile parameters. These features effectively reduce the probability of non-cooperative receivers detecting and identifying their transmitted signals, offering advantages such as high concealment and strong anti-jamming capabilities, and are widely used in modern radar equipment. Currently, various LPI radar signal modulation methods exist, mainly divided into frequency modulation (FM) and phase modulation (PM). FM primarily includes linear frequency modulation (LFM) signals and frequency hopping signals (Costas); PM primarily includes signals phase-coded using Frank codes, Baker codes, P codes (P1-P4), and T codes (T1-T4). The characteristics of LPI radars present significant challenges to the detection and modulation identification of non-cooperative radar signals.

[0003] In non-cooperative radar signal detection tasks, statistical methods were often used in the early stages. First, handcrafted features were designed based on prior knowledge to extract features from the signals detected in the processing frame. Then, a pre-designed decision threshold was used for signal detection. However, traditional methods are ill-suited to increasingly complex electromagnetic environments and the low signal-to-noise ratio, envelope distortion, and low power characteristics of new radar waveforms due to low intercept probability. Furthermore, traditional methods utilize information within a single processing frame, neglecting the sequential relationships between adjacent radar pulses. Accumulating information from multiple frames can improve detection performance under low signal-to-noise ratio conditions.

[0004] In radar signal modulation type identification tasks, traditional methods often combine time-frequency analysis with deep networks. First, time-frequency analysis is performed on the received waveform signal to obtain a two-dimensional time-frequency image. Then, a deep network is used to extract features from the time-frequency image and perform classification. Deep learning-based methods effectively improve the accuracy of modulation type identification under low signal-to-noise ratio conditions. However, time-frequency analysis has high computational complexity, which currently makes it difficult to meet the real-time processing requirements of practical electronic reconnaissance applications.

[0005] Furthermore, in the traditional processing flow of electronic reconnaissance receivers, signal detection and modulation type identification are often handled through two cascaded processes. The first stage performs signal detection, and the second stage identifies the modulation type of the detected signal. However, due to the increasingly complex pulse modulation types of advanced radar systems, on the one hand, different modulation types correspond to different energy distribution patterns in the time and frequency domains, and fully utilizing this modulation information can assist in radar signal detection; on the other hand, cascading the signal detection and modulation identification tasks increases the complexity of the electronic reconnaissance system's processing flow, which is detrimental to practical deployment and application.

[0006] Multi-task learning in deep learning utilizes deep neural networks to solve multiple tasks simultaneously, leading to their widespread application in numerous fields. Multi-task networks can accomplish multiple tasks within a single network, increasing inference speed and expanding their applicability in real-world systems. Furthermore, they can leverage information between different tasks to enhance network performance. In signal detection and modulation recognition tasks, for example, the modulation information of radar signals can assist in signal detection, improving performance under low signal-to-noise ratio conditions. Additionally, a single model solving multiple tasks simultaneously effectively reduces the memory footprint of network parameters and the complexity of deploying in practical systems. Summary of the Invention

[0007] This invention proposes a joint signal detection and modulation identification processing method for low probability of intercept (LPI) radar. It can accurately detect each pulse of the received LPI radar waveform signal containing multiple pulses, complex modulation types, and flexible modulation parameter values, and provide the start and end times and modulation type identification results for each detected pulse.

[0008] A joint signal detection and modulation identification processing method for low probability of intercept (LPI) radars includes:

[0009] S1. Construct the dataset for training:

[0010] S11. Combine multiple pulse waveform samples containing different modulation types into a pulse sequence sample according to their arrival time order;

[0011] S12. Obtain multiple pulse sequence samples with different numbers of pulses, multiple modulation types, and different arrival times, and obtain the arrival time and end time of each pulse in the sample as detection labels, and obtain the modulation type of each pulse in the sample as identification labels; the pulse sequence samples and labels together constitute the dataset.

[0012] S2. Divide the pulse sequence samples in the dataset into multiple sample blocks;

[0013] S3. Construct a deep multi-task neural network JDMR-Net, which includes a local feature extraction layer, a global similarity mining layer, and a task-specific layer.

[0014] The local feature extraction layer is implemented by two LSTM layers oriented towards different time scales. The first LSTM layer performs feature extraction on the preprocessed sample blocks in parallel to obtain the feature extraction results. Feature extraction results Multiple adjacent sample blocks are concatenated to obtain the feature extraction result. Then, the second LSTM layer concatenates the sample blocks. Feature extraction is performed in parallel to obtain the feature extraction results.

[0015] The global similarity mining layer employs a cross-attention mechanism, targeting the feature extraction results from two LSTM layers at different time scales. and feature extraction results Calculate cross-attention to obtain the output of the global similarity mining layer;

[0016] The specific task layer includes a signal detection task layer and a modulation recognition task layer; based on the output of the global similarity mining layer, the signal detection layer outputs the start and end times of each pulse, and the modulation recognition task layer outputs the modulation type corresponding to each pulse; thereby, the deep multi-task neural network JDMR-Net is trained;

[0017] S4. Input the low probability of intercept radar pulse sequence waveform signals to be detected and identified into the corresponding trained deep multi-task neural network JDMR-Net to obtain the start and end times of each pulse in each signal to be detected and identified and its corresponding modulation type.

[0018] Preferably, during the training of the deep multi-task neural network JDMR-Net, the loss function is:

[0019]

[0020] Where α represents the loss function The weight it occupies is 0.1;

[0021]

[0022]

[0023] In the formula, y represents the actual label. b represents the predicted label; 'b' represents the detected label. c represents the predicted detection label, and c represents the recognition label. This represents the predicted identification label, and M represents the number of pulses in the sample. This represents the start time of the j-th pulse. This is the corresponding predicted value output by the signal detection layer; This indicates the end time of the j-th pulse. To modulate the predicted value corresponding to the output of the recognition task layer, c j This indicates the modulation type of the j-th pulse. This indicates the probability of identifying it as this modulation type.

[0024] Preferably, in the global similarity mining layer, the cross-attention mechanism sequentially includes two sets of cascaded attention modules, a feedforward layer, a residual connection layer, and a normalization layer; in the first attention module, As input for generating the query vector q, As input to generate the key vector k and value vector v; in the second attention module, As input for generating the query vector q, As input to generate the key vector k and value vector v;

[0025] Then, for each group of attention modules:

[0026] First, set the query vector q of the current sample block i Key vectors of other sample blocks Dot product is used to calculate the score:

[0027]

[0028] Then divide the score by And through a softmax layer, all scores are normalized:

[0029]

[0030] d k The dimension of the weight matrix that generates the key vectors, where n represents the number of sample blocks;

[0031] Finally, the value vector v of the current sample block i Multiply by the fraction s i ′ j And sum them:

[0032]

[0033] The score z corresponding to each sample block i The fractional vector Z is fed into the feedforward layer, and the output is then processed by the residual connection layer and the normalization layer to obtain the output of the global similarity mining layer.

[0034] Preferably, in step S2, the mean and variance of each pulse sequence sample are first calculated, and the mean is subtracted from each sample and the variance is divided to complete the standardization process, and then the sample is divided into sample blocks.

[0035] Ideally, the detection label should be normalized and the identification label should be one-hot encoded.

[0036] Ideally, cross-attention should be calculated using matrix operations:

[0037]

[0038] Where Q, K, and V are operation matrices composed of multiple query vectors q, key vectors k, and value vectors v, respectively.

[0039] Preferably, the signal detection task layer is implemented by a fully connected layer using the Sigmoid activation function; the modulation type recognition task layer is implemented by a fully connected layer using the Softmax activation function.

[0040] Ideally, the training of the deep multi-task neural network JDMR-Net utilizes the backpropagation algorithm to update the model parameters, repeating the above operation until the maximum number of iterations is reached, while the optimizer Adam is used to reduce the loss function value during training.

[0041] The beneficial effects of this invention are:

[0042] This invention proposes a method for joint signal detection and modulation identification processing of low probability of intercept (LPI) radar signals. This method can accurately detect each pulse in a received LPI radar waveform signal containing multiple pulses, complex modulation types, and flexible modulation parameter values. It also provides the start and end times and modulation type identification results for each detected pulse. Specifically:

[0043] This invention combines the principle of multi-task learning with the automatic feature learning capability of deep neural networks to effectively extract local temporal features and global similarity features from the input original pulse sequence. It can accurately complete signal detection and modulation identification tasks even under poor signal-to-noise ratio conditions. This invention uses waveform sequences containing multiple pulses as input, leveraging the accumulation of multiple pulses to improve detection and identification performance without the additional computational burden of time-frequency transformation in traditional methods. This invention uses a single deep multi-task neural network to simultaneously complete both signal detection and modulation type identification tasks, effectively simplifying the processing flow of traditional electronic reconnaissance receivers. The proposed method for joint radar signal detection and modulation identification processing features low computational complexity and high parallelism, making it possible to deploy the method in practical real-time systems. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the signal processing flow of an electronic reconnaissance receiver.

[0045] Figure 2 This is a schematic diagram of the instantaneous frequency and instantaneous phase of 11 low probability of intercept radar signals simulated in this invention;

[0046] Figure 3 This is a schematic diagram of the low probability of intercept radar pulse sequence waveform signal and the time-frequency converted image simulated by this invention;

[0047] Figure 4 This is a schematic diagram of the detection task and modulation type recognition task corresponding to the multi-task network constructed in this invention;

[0048] Figure 5 This is a hierarchical structure diagram of the joint signal detection and modulation recognition network constructed in this invention;

[0049] Figure 6 This is a hierarchical diagram of the cross-attention mechanism;

[0050] Figure 7 This is a schematic diagram illustrating the parameter ranges of 11 low probability of intercept radar signals simulated in this invention. Detailed Implementation

[0051] This invention provides a method for joint signal detection and modulation identification processing of low probability of intercept radar, comprising the following steps:

[0052] S1. Construct the dataset for training:

[0053] S11. Set the maximum number of pulses that a pulse sequence sample may contain to M. Select multiple modulation types from a set of 11 low probability of intercept (LPI) radar signal modulation types and select the corresponding modulation parameters to obtain a pulse waveform sample. Repeat this process m times (m ≤ M) to obtain m pulse waveform samples. The instantaneous frequency and instantaneous phase of the 11 LP radar signals are as follows: Figure 2 As shown. The m pulse waveform samples containing different modulation types are arranged according to a certain arrival time to form a pulse sequence sample, denoted as . Where T represents the sample length, and 2 indicates that the sample contains both real and imaginary parts. The waveforms and time-frequency diagrams of the pulse sequence under different signal-to-noise ratios are shown below. Figure 3 As shown.

[0054] S12. Obtain multiple pulse sequence samples with different pulse numbers, complex modulation types, and different arrival times, and obtain the arrival and end times of each pulse in the samples as labels for the detection task, such as... Figure 4 As shown, the detection label is denoted as:

[0055] y D=b=(b s ,b e )

[0056] Among them, b s Indicates the starting position of the pulse, b e Indicates the end position of the pulse; obtains the modulation type of each pulse in the sample as an identification tag, and denotes the identification tag as:

[0057] y R =c

[0058] Here, 'c' represents the modulation type; pulse sequence samples and labels together constitute the dataset. The dataset is divided into training, validation, and test sets according to a fixed ratio.

[0059] S2. Sample preprocessing: First, calculate the mean and variance of each sample, then subtract the mean from each sample and divide by the variance to complete the standardization process.

[0060]

[0061] Then, by selecting appropriate segmentation and overlap lengths, the pulse sequence sample is divided into multiple small sample blocks, which facilitates subsequent parallel processing of these small sample blocks.

[0062]

[0063] Among them, X · represents a small sample block, n represents the number of sample blocks, and w represents the size of the sample block.

[0064] Preprocessing of the label samples: Normalize the detection labels and perform one-hot encoding on the identification labels.

[0065] S3. Construct a joint detection and modulation recognition multi-task deep neural network JDMR-Net: such as Figure 5 As shown, JDMR-Net includes a local feature extraction layer, a global similarity mining layer, and detection and recognition task layers. The local feature extraction layer is implemented by two consecutive LSTM layers. The first LSTM layer is oriented at a time scale of a single small sample block, and the second LSTM layer is oriented at a time scale of multiple adjacent small sample blocks. The global similarity mining layer is implemented using a cross-attention mechanism, calculating cross-attention features for the features extracted by the two LSTM layers at different time scales. Finally, the detection task layer and the recognition task layer receive the extracted deep features and output the start and end times of the pulses and the corresponding modulation types, respectively. Specifically:

[0066] S31, Local Feature Extraction Layer

[0067] The local feature extraction layer is implemented by two LSTM layers targeting different time scales, enabling effective extraction of intra-pulse modulation features at different time scales. First, the first LSTM layer processes the preprocessed signal... Feature extraction is performed in parallel on n small sample blocks. Here, the time scale of LSTM feature extraction is one sample block length. The LSTM layer contains 256 hidden nodes and uses the ReLU activation function to obtain the feature extraction results.

[0068]

[0069] The n adjacent features in the first layer feature extraction result l By splicing together small sample blocks, we get:

[0070]

[0071] Then, the second LSTM layer is spliced ​​together. Feature extraction is performed in parallel on small sample blocks, where the time scale of LSTM feature extraction is n. l The sample block length is given, the LSTM layer contains 256 hidden nodes, and the ReLU activation function is used to obtain the feature extraction results:

[0072]

[0073] S32, Global Similarity Mining Layer

[0074] The global similarity mining layer relies on a cross-attention mechanism to target different features extracted by two LSTM layers at different time scales. and The cross-attention mechanism is calculated by sequentially connecting the attention layer, feedforward layer, residual connection, and normalization layer, as follows: Figure 6 As shown. When calculating cross-attention, in the first cross-attention module, As input for generating the query vector q, As input to generate the key vector k and value vector v; in the second cross-attention module, the opposite operation is performed. As input for generating the query vector q, As input to generate the key vector k and value vector v, the first step in calculating cross-attention is to compute the query vector q, key vector k, and value vector v, which are obtained by multiplying the inputs by their respective feature matrices:

[0075] q = W Q

[0076] k = X·W K

[0077] v = W V

[0078] and These are three trainable weight matrices; X represents the data input to the two cross-attention modules, respectively. and d q ,d k and d v The dimensions of the weight matrix are respectively

[0079] The second step in calculating cross-attention is to calculate a score, which measures the similarity of the current sample block with other samples. The score is calculated using the query vector q of the current sample block. i Key vectors of other sample blocks The result obtained by taking the dot product:

[0080]

[0081] The third step in calculating cross-attention is to divide the score by... Then, by using a softmax layer, the scores of all words are normalized, resulting in scores that are all positive and sum to 1.

[0082]

[0083] d k The dimension of the weight matrix that generates the key vectors;

[0084] The fourth step in calculating cross-attention is for each value vector v i The scores obtained by multiplying by the softmax function and then summing them are:

[0085]

[0086] To achieve fast computation, the above steps for calculating cross-attention are often performed in the form of matrix operations:

[0087]

[0088] Where Q, K, and V are operation matrices composed of multiple query vectors q, key vectors k, and value vectors v, respectively.

[0089]

[0090] Then, Z is fed into the feedforward layer for feature extraction, which includes two linear transformation layers using the ReLU activation function, denoted as,

[0091] FC(z) = max(0, W1z+1)W2+2

[0092] Where z represents the row vector in Z, and W1 and W2 are two trainable parameter matrices.

[0093] Then, residual connection and normalization operations are performed on the output of the feedforward layer, LayerNorm(+Sublayer()), where Sublayer(·) represents the output of the self-attention layer or the feedforward layer, to obtain the output of the global similarity mining layer.

[0094] S33, Specific Task Layer

[0095] The specific task layer includes a signal detection task output layer and a modulation recognition task output layer. The signal detection task layer is implemented by a fully connected layer, using the sigmoid activation function, and outputs the normalized start and end times corresponding to each detected pulse; the modulation type recognition task layer is implemented by a fully connected layer, using the softmax activation function, and outputs the probability that each detected pulse belongs to each modulation type.

[0096] S4. Train the JDMR-Net model

[0097] S41. Constructing the loss function

[0098]

[0099]

[0100]

[0101] In the formula, y represents the actual label. Let b represent the predicted label, c represent the detected label, and M represent the number of pulses in the sample. This represents the start time of the j-th pulse. c represents the end time of the j-th pulse. j This indicates the modulation type of the j-th pulse. This indicates the probability of identifying it as this modulation type.

[0102] S42. Train the JDMR-Net model using the training and validation sets, and save the JDMR-Net model after training is stable. During model training, first initialize the model parameters, then calculate the joint loss function value through forward propagation, and finally update the model parameters using the backpropagation algorithm. Repeat the above operations until the maximum number of iterations. At the same time, use the optimizer Adam to reduce the loss function value during training, so that the predicted distribution and the target distribution gradually approach each other, thereby improving the performance of the model.

[0103] S5. After preprocessing (using the method in S2), the low probability of intercept radar pulse sequence waveform signals to be detected and identified are input into the corresponding trained deep multi-task neural network JDMR-Net to obtain the start and end times and modulation type of each detected pulse.

[0104] Example:

[0105] S1. First, generate a sample dataset for model training using real or simulated data:

[0106] S11. Based on relevant domain expert knowledge, clean, extract, or simulate to generate 14 datasets D1 with different signal-to-noise ratios (SNRs). D1 contains 14 data subsets, corresponding to 14 different SNRs: [-16, -14, -12, -10, -8, -6, -4, -2, 0, 2, 4, 6, 8, 10]. Each sample in the data subset is a pulse waveform sequence, denoted as... A single sample can contain a maximum of 20 pulses and 2 different intra-pulse modulation types. The sampling frequency of the sample is 20MHz, and the pulse width is 5-10µs. The modulation parameter ranges used for the 11 modulation types are as follows: Figure 7 As shown.

[0107] Based on the knowledge of experts in the relevant field, a non-ideal dataset D2 containing envelope distortion is generated through cleaning, extraction, or simulation. D2 can be generated from dataset D1 by adding Hamming windows to the pulses of each sample in dataset D1 to simulate pulse envelope distortion under non-ideal conditions.

[0108] Based on the knowledge of experts in the relevant field, a non-ideal dataset D3 containing envelope distortion and spurious pulses is generated through cleaning, extraction, or simulation. D3 can be generated from dataset D2 by adding simple rectangular pulses to the locations where no pulses appear in each sample of dataset D2 to simulate the pulse situation under non-ideal conditions.

[0109] S12. Obtain the label corresponding to each sample. Each pulse in the sample contains a detection label and a recognition label, denoted as .

[0110]

[0111] The arrival and end times of each pulse are used as detection tags y. D The modulation type of each pulse serves as an identification tag y R For example, pulse waveform sequences containing LFM, Costa, and Frank code modulation types, such as... Figure 3 The labels corresponding to the sequences shown can be represented as

[0112] y={(57,357,582,882,1268,1568,2393,2693),(LFM,Costas,LFM,Frank)}

[0113] S2. Sample preprocessing: First, calculate the mean and variance of each sample, then subtract the mean from each sample and divide by the variance to complete the standardization process.

[0114]

[0115] Then, by selecting appropriate segmentation and overlap lengths, the pulse sequence sample is divided into multiple small sample blocks, which facilitates subsequent parallel processing of these small sample blocks.

[0116]

[0117] Preprocessing of the labeled samples: Normalization of the detection labels and one-hot encoding of the recognition labels to obtain the label sequence.

[0118] y={(0.0190,0.1190,0.1940,0.2940,0.4227,0.5227,0.7977,0.8977),

[0119] (00000000001,00000000010,00000000001,00000000100)}

[0120] S3. Construct a joint detection and modulation recognition multi-task deep neural network JDMR-Net, including a local feature extraction layer, a global similarity mining layer, and detection task layers and modulation recognition task layers, such as... Figure 5 As shown.

[0121] S31, Local Feature Extraction Layer

[0122] The local feature extraction layer is implemented using two LSTM layers oriented towards different time scales. First, the first LSTM layer processes the preprocessed signal... Feature extraction is performed in parallel on n small sample blocks. The LSTM layer contains 256 hidden nodes and uses the ReLU activation function to obtain the features. Will n adjacent l Small sample blocks are spliced ​​together to obtain The second LSTM layer performs feature extraction in parallel on the concatenated small sample blocks. Here, the time scale for LSMT feature extraction is n. l The sample block length is given, the LSTM layer contains 256 hidden nodes, and the ReLU activation function is used to obtain the features.

[0123] S32, Global Similarity Mining Layer

[0124] Using input vectors and Three vectors are generated: query vector q, key vector k, and value vector v.

[0125] Then calculate the self-attention score s. ij s is obtained through the Softmax layer i ′ j ;

[0126] The score obtained by multiplying each value vector by Softmax is then calculated as follows:

[0127]

[0128] To enable fast computation, the steps for calculating self-attention are often performed in matrix form, such as... Figure 6 As shown

[0129]

[0130] Then, Z is fed into the feedforward layer for feature extraction, and the output of the feedforward layer is subjected to residual connection and normalization operations, LayerNorm(+Sublayer()), where Sublayer(·) represents the output of the attention layer or the feedforward layer, to obtain the output of the global similarity mining module.

[0131] S33, Specific Task Layer

[0132] The specific task layer comprises a signal detection task layer and a modulation recognition task layer. The signal detection task layer is implemented using a fully connected layer and a sigmoid activation function, outputting the normalized start and end times of each pulse. The modulation recognition task layer is also implemented using a fully connected layer and a softmax activation function, outputting the probability that each pulse belongs to each modulation type.

[0133] S4. Utilize the real label y and the network output label. Calculate the loss function:

[0134]

[0135]

[0136]

[0137] By using the optimizer Adam to reduce the loss function value during training, the predicted distribution and the target distribution gradually approach each other, resulting in a well-trained JDMR-Net network.

[0138] S5. Input the low probability of intercept radar pulse sequence waveform signal to be detected and identified into the JDMR-Net network trained in step S4 to obtain the start and end times and modulation type of each pulse in the sample.

[0139] In summary, the above are merely implementation examples of this invention based on selected radar pulses and 11 LPI intra-pulse modulation types, and are not intended to limit the scope of protection of this invention. The core elements of this invention are: a segmentation-based data preprocessing method, a local feature extraction module for different time scales, a global similarity mining module relying on cross-attention mechanisms, and a multi-task network for joint detection and modulation recognition utilizing accumulated information from multiple pulses. Any low-probability-of-interception radar signal joint signal detection and modulation recognition processing method formed by corresponding modifications, substitutions, and improvements within the above design principles and implementation points of this invention should be included within the scope of protection of this invention.

Claims

1. A method for joint signal detection and modulation identification of low probability of intercept radars, characterized in that, Comprise: S1, constructing a data set for training: S11, a plurality of pulse waveform samples containing different modulation types are grouped into a pulse sequence sample in the order of arrival time; S12, a plurality of pulse sequence samples with different pulse numbers, modulation types and different arrival times are obtained, and the arrival time and end time of each pulse in the sample are obtained as detection labels, and the modulation type of each pulse in the sample is obtained as identification labels; The pulse sequence sample and the label jointly constitute a data set; S2, the pulse sequence samples in the data set are divided into a plurality of sample blocks; S3, constructing a deep multi-task neural network JDMR-Net, which includes a local feature extraction layer, a global similarity mining layer and a specific task layer; The local feature extraction layer is implemented by two LSTM layers facing different time scales, a first LSTM layer performs feature extraction on the preprocessed sample blocks in parallel to obtain a feature extraction result The feature extraction result is spliced with a plurality of adjacent sample blocks to obtain a feature extraction result Then, a second LSTM layer performs feature extraction on the spliced sample blocks in parallel to obtain a feature extraction result The global similarity mining layer is implemented by using a cross-attention mechanism, and is used for extracting feature extraction results of features extracted by two LSTM layers facing different time scales and feature extraction results Calculate cross-attention to obtain an output of the global similarity mining layer. The specific task layer includes a signal detection task layer and a modulation identification task layer; based on the output of the global similarity mining layer, the signal detection layer outputs the start and end time of each pulse, and the modulation identification task layer outputs the modulation type corresponding to each pulse; thereby training the deep multi-task neural network JDMR-Net; S4, the low probability of interception radar pulse sequence waveform signal to be detected and identified is input into the corresponding trained deep multi-task neural network JDMR-Net, and the start and end time of each pulse in each signal to be detected and identified and the corresponding modulation type are obtained.

2. The method for joint signal detection and modulation identification of low probability of intercept radars according to claim 1, characterized in that, In the process of training the deep multi-task neural network JDMR-Net, the loss function is: wherein a represents the loss function The weight occupied is 0.

1. where y denotes the true label, denotes the predicted label; b denotes the detected label, denotes the predicted detected label, c denotes the recognized label, denotes the predicted recognized label, M denotes the number of pulses in the sample, denotes the start time of the jth pulse, is the corresponding predicted value output by the signal detection layer; denotes the end time of the jth pulse, is the corresponding predicted value output by the modulation recognition task layer, c j denotes the modulation type of the jth pulse, denotes the probability of deciding the modulation type.

3. The method of joint signal detection and modulation identification of low probability of intercept radars according to claim 1 or 2, characterized in that, In the global similarity mining layer, the cross attention mechanism comprises two groups of attention modules, a feedforward layer, a residual connection layer and a normalization layer in sequence; in the first attention module, as input for generating the query vector q, as input for generating the key vector k and the value vector v; in the second attention module, as input for generating the query vector q, as input for generating the key vector k and the value vector v; Then, each group of attention modules: First, the query vector q of the current sample block is dot producted with the key vectors of other sample blocks i and other sample blocks' key vectors Dot product is done to calculate the score: The scores are then divided by and normalized by a softmax layer, making all scores sum to one: d k For the dimension of the weight matrix that produces the key vector, n represents the number of sample blocks. Finally, the value vector v of the current sample block is multiplied by the fraction s i i ′ j and summed up:​ The score z corresponding to each sample block i The score vector Z is fed into the feedforward layer, and the output is further processed by the residual connection layer and the normalization layer to obtain the output of the global similarity mining layer.

4. The method for joint signal detection and modulation identification of low probability of intercept radar of claim 1, wherein, In S2, the mean and variance of each pulse sequence sample are calculated, and each sample is subtracted by the mean and divided by the variance to complete the standardization processing, and then divided into sample blocks.

5. The method for joint signal detection and modulation identification of low probability of intercept radar of claim 1, wherein, The detection label is normalized, and the identification label is one-hot encoded.

6. The method for joint signal detection and modulation identification of low probability of intercept radar of claim 1, wherein, The cross-attention is calculated in the form of matrix operation: Wherein, Q, K, V are operation matrices composed of a plurality of query vectors q, key vectors k and value vectors v.

7. The method for joint signal detection and modulation identification of low probability of intercept radar of claim 1, wherein, The signal detection task layer is realized by a fully connected layer, and uses a Sigmoid activation function; the modulation type identification task layer is realized by a fully connected layer, and uses a Softmax activation function.

8. The method for joint signal detection and modulation identification of low probability of intercept radar of claim 1, wherein, The training of the deep multi-task neural network JDMR-Net updates the model parameters by using the back propagation algorithm, and repeatedly performs the above operations until the maximum iteration number, while using the optimizer Adam to reduce the loss function value in the training process.

Images