A deep learning radar signal deinterleaving method based on differential attention assistance

Abstract

The application discloses a kind of deep learning radar signal deinterleaving methods based on differential attention auxiliary, it is related to radar technical field, the application includes, S1. The TOA sequence of received signal is generated and is different, and DTOA sequence is obtained;S2. Join missing pulse and false pulse, obtain data set;S3. Design differential attention mechanism, construct differential attention module assisted deep neural network model;S4. Using data set trains and evaluates model, obtains the best neural network model to carry out deinterleaving to the DTOA sequence of aliasing.The application is combined organically by differential attention mechanism and deep learning network, under the premise of not depending on artificial feature extraction, high-precision, high robustness deinterleaving to radar signal is realized, especially suitable for high interference, multiple modulation, complex aliasing actual battlefield environment, provides effective technical support for radar electronic reconnaissance and countermeasure.

Description

Technical Field

[0001] This invention relates to the field of radar technology, and specifically to a deep learning-based radar signal deinterleaving method based on differential attention assistance. Background Technology

[0002] Radar signal sorting is one of the key technologies in radar electronic reconnaissance. Firstly, when detecting targets, effective signal sorting can minimize the impact of interference signals and false targets on real targets, thereby improving the accuracy of target identification. Secondly, in complex electromagnetic environments, only accurate signal sorting can effectively identify the properties of radiation sources and assess threats, enabling targeted countermeasures to be taken accordingly. Therefore, radar signal sorting technology has attracted widespread attention.

[0003] Radar signal sorting refers to the process of deinterleaving received signals from different radiation source targets that are mixed in the time domain, based on their signal differences, to obtain different PDW sequences. Existing research mainly falls into two categories: traditional methods that use thresholds to estimate potential PRI, which usually require manual extraction of inter-pulse features and are difficult to handle complex modulation situations; and deep learning methods based on neural networks, which usually require preprocessing of TOA sequences, and have the disadvantages of not being able to comprehensively consider the original information and having poor robustness in high-interference environments.

[0004] For the first type of method, the paper "An overview and classification of machine learning approaches for radar signal deinterleaving" proposes a typical method based on the statistical histogram of pulse repetition intervals. This method manually sets a threshold to estimate potential PRIs and then repeats the search in the original sequence. However, its performance degrades in environments where signals are lost. To address this issue, the paper "Radar signal binning based on improved SDIF algorithm" proposes a typical method based on sequence difference histograms. This method analyzes continuous difference patterns in the pulse sequence by calculating histograms of different orders, thereby identifying the variation patterns of PRIs. However, this method struggles to deinterleave PRI signals with irregular variations. To address this issue, the paper "A radar signal sorting algorithm based on improved PRI transform" proposes a typical method based on the PRI transform. This method maps the time-domain PRI sequence to the transform domain for multi-dimensional feature analysis. However, the selection of transform parameters requires manual intervention, leading to unstable deinterleaving performance.

[0005] For the second type of method, the paper "Classification, denoising, and deinterleaving of pulse streams with recurrent neural networks" proposes a typical method based on recurrent neural networks. This method uses multiple RNNs for binary classification prediction. However, using multiple pre-trained networks can lead to computational complexity. To address this issue, the paper "Image segmentation for radar signal deinterleaving using deep learning" proposes a typical method based on U-Net. This method preprocesses pulse sequences to obtain radar pulse images. However, the preprocessed pulse images lose original information, and the simple U-Net network struggles to extract features between long-range signals. To address this issue, the paper "A radar signal deinterleaving method based on semantic segmentation with neural network" proposes a typical method based on bidirectional long short-term memory networks and dilated convolutional networks, capable of extracting features from long-range signals. However, it still cannot comprehensively consider original information across multiple dimensions, and its deinterleaving accuracy is unstable in complex environments with high interference. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a deep learning-based radar signal deinterleaving method based on differential attention assistance to improve the deinterleaving performance and anti-interference performance of radar pulse signals.

[0007] In a first aspect, the present invention provides a deep learning-based radar signal deinterleaving method based on differential attention assistance, such as... Figure 1 As shown, it includes the following steps:

[0008] S1. Generate the TOA sequence of the received signal and perform differential processing to obtain the DTOA sequence, i.e., the PRI modulation sequence;

[0009] S2. To simulate a real battlefield environment, a certain proportion of lost pulses and false pulses are added to obtain a dataset;

[0010] S3. Design a differential attention mechanism and construct a deep neural network model assisted by a differential attention module;

[0011] S4. Train and evaluate the model using the dataset to obtain the optimal neural network model for deinterleaving aliased DTOA sequences.

[0012] Furthermore, step S1 includes:

[0013] Three TOA sequences with common PRI modulation characteristics are generated, aliased, and differencing to obtain multiple sets of PRI-modulated time series sequences of length n. Each PRI sequence corresponds to a label, which is a 1×n vector. Specifically: Three TOA sequences with common PRI modulation characteristics are generated, aliased, and differencing to obtain 10,000 sets of DTOA (i.e., PRI-modulated) time series sequences of length 1000. Each PRI sequence corresponds to a label, which is a 1×1000 vector. The three common PRI modulation methods are as follows:

[0014] (1) Constant PRI

[0015] Many radars generate pulse trains using a fixed pulse repetition interval. However, depending on the stability of the employing device, the pulse repetition period may drift slowly over time. This type of signal is commonly used in search and tracking radars, especially those where the PRI (Pulse Repetition Period) remains constant. The PRI sequence can be represented as...

[0016]

[0017]

[0018] In the formula: Representing the first One pulse, It represents a constant.

[0019] (2) Dwell & Switch (D&S) PRI modulation

[0020] The PRI period of the radar pulse stream in this modulation mode is determined by the number of PRI groups. Each value exists for a duration and is then immediately replaced by a subsequent value. The PRI values ​​change periodically between groups. The PRI sequence can be represented as...

[0021]

[0022]

[0023] For the first in each group , This indicates the number of pulses in each group. This represents the number of pulse groups within one cycle, i.e., the number of pulse groups within one cycle. The number of values.

[0024] (3) Uneven PRI

[0025] The pulse sequence modulated by this mode is a radar pulse sequence with two or more different PRI values, which appear periodically. The mathematical representation of the PRI variation in staggered PRI modulation is as follows:

[0026]

[0027] Where M represents the number of PRI values ​​within a period.

[0028] Furthermore, step S2 includes:

[0029] For the parameter settings of lost pulses and false pulses, the loss rate of the target pulse is denoted as... The ratio of the number of random noise pulses in the intercepted pulse stream to the average number of target radar pulses is used. This represents the impulse noise rate. Since the impulse loss rate and impulse noise have a significant impact on the impulse sequence, the range is set to 0 < 0. < 0.25 and 0 < < 0.25.

[0030] Furthermore, step S3 includes:

[0031] S301. A differential attention mechanism is designed based on the characteristics of PRI sequences, directly focusing on the temporal difference characteristics of the signal. It is particularly suitable for detecting the PRI mode in DTOA sequences and has a significant enhancement effect on signal deinterleaving in aliased signals.

[0032] For the differential attention mechanism, the architecture is as follows: Figure 3 As shown, the specific implementation is as follows:

[0033] The arrival time is differentially analyzed to obtain the PRI parameters. By determining the PRI modulation scheme, signals from different transmitters are separated from the interleaved signals. The differential attention mechanism combines standard attention calculation with differential feature modeling. Standard attention calculates the similarity between features through query-key-value (QKV) projection. The calculation process of the standard attention score is shown in the following formula:

[0034]

[0035] in, It is the input feature representation. These are the linear projection parameters of the query and the key. It is a query and key representation. It is an attention score matrix. It is a standardization factor to ensure gradient stability;

[0036] Differential attention explicitly computes temporal difference features to capture similarity patterns in pulse intervals. The difference features are obtained by subtracting features from those of adjacent time steps. After feature encoding, a similarity matrix is ​​calculated. The calculation process is shown below:

[0037]

[0038]

[0039] It is the difference vector between adjacent time steps. It is the parameter matrix of the differential feature encoder. Representation layer normalization operations ensure training stability. It is the SiLU activation function, where It is the sigmoid function. These are the encoded differential features. It is a difference similarity matrix. It is a normalization factor that ensures the stability of the dot product value;

[0040] Finally, the standard attention score and the difference similarity matrix are weighted and fused, then normalized using Softmax for weighted feature aggregation. Standard attention and difference attention complement each other in information extraction: standard attention focuses on absolute feature similarity, while difference attention focuses on time interval patterns. The sequence dependencies captured by LSTM are further enhanced by difference attention, and both difference features and sequence features are utilized, making the model highly sensitive to complex patterns in aliased DTOA sequences.

[0041] S302. A deep neural network is constructed based on differential attention assistance. This method can fully utilize the original inter-pulse features to adapt to complex modulation environments. A deep learning network based on differential attention (DA) assistance (DA-DL) has the following structure: Figure 2 As shown;

[0042] Features of DTOA sequences are extracted using a multi-scale feature extraction network. Then, a differential attention mechanism is used to enhance the extraction of semantic features of PRI modulation mode after temporal modeling. Finally, the outputs of temporal modeling and differential attention are fused for signal classification.

[0043] (1) Multi-scale feature extraction network

[0044] Design a multi-scale feature extraction network to extract sequence features at multiple scales. The multi-scale feature extraction network is cascaded with a multi-scale feature fusion (MSFF) module and a dilated convolutional network (DCN).

[0045] First, the MSFE module uses convolutional layers with different receptive fields in parallel, enabling the extraction of features at different time scales. Then, a DCN is used for deep feature extraction. To maintain the feature map length constant in each convolutional step, the DCN does not use causal convolutions but instead symmetrically fills both sides of the feature map. This ensures that the receptive field of each convolutional kernel in the last layer of the network is sufficient to cover the length of the input data. In this scheme, the DCN's convolutional kernel size is set to 5, and four residual blocks are cascaded to exponentially expand the receptive field, effectively capturing long-range dependencies. The MSFE network achieves comprehensive feature extraction of DTOA sequences, capturing both local detail patterns and understanding long-term dependencies. This provides rich feature representations for subsequent temporal modeling and classification.

[0046] Traditional Inception modules capture multi-scale features of images by deploying convolutional paths of different scales in parallel, but their structure is complex and computationally expensive. To address the characteristics of the time-series data studied in this invention, an improved MSFF module is proposed. This module removes the 1×1 convolutional dimensionality reduction operation and pooling layers, directly using convolutional kernels of different scales for feature extraction. It employs three parallel convolutional paths, using convolutional layers with kernel sizes of 3, 5, and 9 respectively to capture local features, intermediate time-scale features, and long-term dependencies of the input sequence. The workflow is as follows:

[0047]

[0048] in, denoted by the scale factor, with values ​​of 1, 2, and 4 respectively; p is the padding size; k is the kernel size; x is the input data; Conv1d represents one-dimensional convolution; GroupNorm is used for feature normalization; and GELU is used as the activation function. This improvement allows the model to maintain stable performance during mini-batch training. Furthermore, using GELU instead of ReLU provides smoother gradients, which helps the model converge during training.

[0049] In the feature fusion stage, the multi-scale feature fusion module concatenates features from three scales along the channel dimension:

[0050] Subsequently, feature fusion is performed using 1×1 convolution to generate the final multi-scale feature representation:

[0051]

[0052] Next, a dilated convolutional network (DCN) is used for deep feature extraction. The DCN shares the same residual blocks as the temporal convolutional network, and the DCN symmetrically fills both sides of the feature map. To maintain the length of the feature map in each convolutional step, the DCN does not use causal convolution but instead symmetrically fills both sides of the feature map. This ensures that the receptive field of each convolutional kernel in the last layer of the network is sufficient to cover the length of the input data. In this invention, the DCN's convolutional kernel size is set to 5, and four residual blocks are cascaded to exponentially expand the receptive field, thereby effectively capturing long-range dependencies. Each dilated residual block contains two weight-normalized dilated convolutional layers, a SiLU activation function, and a regularization layer, and uses SE channel attention to adaptively adjust channel importance and enhance key features.

[0053] (2) Temporal modeling module

[0054] After deep feature extraction from multi-scale sequences, the process also includes reconstructing the time series:

[0055] After deep feature extraction from multi-scale sequences, time series reconstruction is required. The Temporal Modeling (TM) module is used to capture long-term and short-term dependencies in the signal sequence, enhancing the modeling capabilities of time series. For DTOA data, forward and backward information are equivalent. To fully utilize the complete information of the pulse sequence to determine pulse categories, positional encoding is introduced into the TM, providing positional information for each time step in the sequence. A Bidirectional Long Short-Term Memory (BLSTM) network is then used to capture forward and backward dependencies. Positional encoding is implemented using sine and cosine functions. Positional encoding using sine and cosine functions requires no parameter learning and can effectively represent the relative relationships between different positions.

[0056] For the dataset obtained in S2, which includes a training set and a validation set, a deep neural network model is trained using the training set and evaluated using the validation set. The model parameters are continuously adjusted to obtain the optimal neural network model. This network is then used to de-interleave the aliased DTOA sequences, and the dataset can be divided into a training set and a validation set in a 9:1 ratio.

[0057] This invention provides a deep learning-based radar signal deinterleaving method based on differential attention assistance. Recognizing the natural match between semantic feature extraction of PRI modulation modes and the characteristics of differential attention mechanisms, this invention proposes a differential attention (DA)-assisted deep learning method. First, a DTOA signal sequence is generated. To simulate a real environment, a certain proportion of missing pulses and spurious pulses are added to the sequence to construct a dataset. Then, to fully utilize the original inter-pulse features to adapt to complex modulation environments, a differential attention mechanism is designed, and a deep network is constructed based on it. Finally, this deep network is used to train the dataset and perform deinterleaving and signal classification.

[0058] This invention achieves high-precision and robust deinterleaving of radar signals by organically combining differential attention mechanism with deep learning network without relying on manual feature extraction. It is particularly suitable for actual battlefield environments with high interference, multiple modulations, and complex aliasing, providing effective technical support for radar electronic reconnaissance and countermeasures. Attached Figure Description

[0059] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and constitute a part of this invention, are not intended to limit the scope of the invention. In the drawings:

[0060] Figure 1 This is a diagram illustrating the signal deinterleaving process.

[0061] Figure 2 This is a diagram illustrating the overall architecture of the method of the present invention;

[0062] Figure 3 This is a diagram of the differential attention architecture of the present invention;

[0063] Figure 4(a) is = Confusion matrix diagram of BLSTM method when ω = 0.2;

[0064] Figure 4(b) is = The confusion matrix diagram of the method of the present invention when = 0.2;

[0065] Figure 5(a) shows the accuracy curve without noise impulse interference;

[0066] Figure 5(b) shows the accuracy curve without pulse loss rate;

[0067] Figure 5(c) shows the accuracy curves with both impulse noise and impulse loss rate. Detailed Implementation

[0068] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention.

[0069] Technical concept of the present invention:

[0070] Radar signal sorting refers to deinterleaving the signals of different radiation source targets received and mixed in the time domain according to their signal differences to obtain different PDW sequences. Existing research mainly falls into two categories: traditional methods that estimate potential PRIs using thresholds, usually requiring manual extraction of inter-pulse features and being difficult to handle complex modulation situations; deep learning methods based on neural networks, usually requiring preprocessing of the TOA sequence and having the disadvantages of not being able to comprehensively consider the original information and having poor robustness in high-interference environments. The present invention proposes a deep learning method assisted by differential attention (DA), which can make full use of the original inter-pulse features to adapt to complex modulation environments. First, a DTOA signal sequence is generated. To simulate the real environment, a certain proportion of lost pulses and false pulses are added to the sequence to construct a dataset. Then, in order to make full use of the original inter-pulse features to adapt to complex modulation environments, a differential attention mechanism is designed and a deep network is constructed based on this. Finally, the dataset is trained using this deep network to deinterleave and classify the signals.

[0071] A deep learning radar signal deinterleaving method based on differential attention assistance provided by the present invention aims to solve the above technical problems of the existing technology.

[0072] The technical solution of the present invention and how the technical solution of the present invention solves the above technical problems will be described in detail below with specific embodiments. These specific embodiments below can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present invention will be described below in conjunction with the drawings.

[0073] Embodiment 1: To verify the proposed scheme, the existing scheme and the proposed scheme are used to deinterleave the signals.

[0074] To verify the proposed scheme, the present invention compares the sorting accuracy rate, mean recall rate (MR), mean precision (MP), and mean intersection over union (MIOU) of BLSTM and the proposed method. The parameter settings are as follows: Three modulation methods, namely constant PRI, switch & dwell (D&S PRI), and interleaved PRI, are adopted to design a feasibility experiment for sorting different PRI modulation methods: For different PRI modulation modes, their PRI values satisfy 20 < PRI < 100; the number of pulses in each group of D&S PRI satisfies 3 J 7, and the number of pulse groups in one cycle satisfies 3 M 7; for interleaved PRI, the number of PRI values in one cycle satisfies the condition 2 M 10; Gaussian distributed variance is added to the TOA to simulate measurement errors, with a standard deviation of 0.1.

[0075] A. Performance metrics under different methods

[0076] Table 1. Performance (%) of each method

[0077] Method ACC MR MP MIOU BLSTM 90.5 74.66 81.84 67.35 DCN 82.30 66.29 71.26 56.28 The proposed method 94.6 87.02 90.10 80.89

[0078] Table 1 shows that the method of this invention outperforms the other two networks in overall performance for the complex PRI modulation mode deinterleaving problem. The proposed method significantly surpasses the other two models in all key metrics, especially in MR and MIOU, where its performance is more than 10% higher than that of the BLSTM method, demonstrating a significant advantage.

[0079] B. Confusion matrix under different methods

[0080] To provide a more detailed analysis of the performance of the proposed method, this invention presents... = =0.2, the confusion matrix of the proposed method is shown in Figure 4, and it is compared with the BLSTM method. Figure 4(a) shows... = The confusion matrix of the BLSTM method when ω = 0.2; Figure 4(b) shows... = The confusion matrix diagram of the method of the present invention when = 0.2;

[0081] The confusion matrix shows that, under conditions of high pulse loss rate and high pulse noise, the sorting accuracy of D&S modulation and interleaved modulation is significantly improved, by approximately 9% and 6%, respectively, indicating that the method of the present invention has a significant advantage in sorting complex PRI modulation modes. Due to the similarity between D&S signals, interleaved signals, and noise pulses, pulse noise is easily confused and classified as DS signals and interleaved signals. The confusion matrix shows that the method of the present invention also improves the sorting accuracy for pulse noise, reducing misclassification.

[0082] C. Accuracy curves under different interference conditions

[0083] Figure 5 shows the classification accuracy changes of the three models (the proposed method, BLSTM, and DCN) with increasing interference, under conditions of no noise impulse interference, no impulse loss rate, and both impulse noise and impulse loss rate. Figure (a) shows the changes in classification accuracy as the interference increases. As the noise level increases, the classification accuracy begins to decline, but the accuracy of the proposed method consistently outperforms the other two methods, demonstrating a higher tolerance for pulse loss. In Figure (b), the accuracy of the three methods does not change significantly, and all models are less sensitive to noise pulses than to pulse loss. Figure (c) shows that under superimposed interference, the classification difficulty increases dramatically, but the proposed method consistently maintains an accuracy superior to the other two methods, remaining above 80%, demonstrating better robustness.

[0084] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0085] Furthermore, in the embodiments of the present invention, the functional modules can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0086] Those skilled in the art will understand that embodiments of the present invention can be provided as methods or systems. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0087] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0088] The above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

[0089] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the invention disclosed herein in the specification and examples. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the foregoing claims.

[0090] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A deep learning-based radar signal deinterleaving method based on differential attention assistance, characterized in that, Includes the following steps: S1. Generate the TOA sequence of the received signal and perform differential analysis to obtain the DTOA sequence; S2. Add missing pulses and spurious pulses to obtain the dataset; S3. Design a differential attention mechanism and construct a deep neural network model assisted by a differential attention module; S4. Train and evaluate the model using the dataset to obtain the optimal neural network model for deinterleaving aliased DTOA sequences.

2. The method for deinterleaving deep learning radar signals based on differential attention assistance according to claim 1, characterized in that, Step S1 includes: Three TOA sequences with common PRI modulation features were generated, aliased, and differentially divided to obtain multiple sets of PRI modulation time sequences of length n. Each PRI corresponds to a label, which is a 1×n vector.

3. The method for deinterleaving deep learning radar signals based on differential attention assistance according to claim 2, characterized in that, Step S2 includes: For the parameter settings of lost pulses and false pulses, the loss rate of the target pulse is denoted as... The ratio of the number of random noise pulses in the intercepted pulse stream to the average number of target radar pulses is used. This represents the impulse noise rate.

4. The method for deinterleaving deep learning radar signals based on differential attention assistance according to claim 3, characterized in that, Step S3 includes: A multi-scale feature extraction network is used to extract features from DTOA sequences. Then, a differential attention mechanism is used to enhance the extraction of semantic features of PRI modulation mode after temporal modeling. Finally, the outputs of temporal modeling and differential attention are fused for signal classification. Among them, a multi-scale feature extraction network is designed to extract sequence features at multiple scales. The multi-scale feature extraction network is cascaded with a multi-scale feature fusion module and a dilated convolutional network. For the multi-scale feature fusion module, the 1×1 convolutional dimensionality reduction operation and pooling layer are removed. Feature extraction is performed directly using convolutional kernels of different scales. Three parallel convolutional paths are used, employing convolutional layers with kernel sizes of 3, 5, and 9 to capture local features, intermediate time-scale features, and long-term dependencies of the input sequence. The process is as follows: in, is the scale factor, with values ​​of 1, 2, and 4 respectively; p is the padding size; k is the kernel size; x is the input data; Conv1d represents one-dimensional convolution; GroupNorm is used for feature normalization; and GELU is used as the activation function. In the feature fusion stage, the multi-scale feature fusion module concatenates features from three scales along the channel dimension: Subsequently, feature fusion is performed using 1×1 convolution to generate the final multi-scale feature representation: Next, a dilated convolutional network is used for deep feature extraction. The DCN shares the same residual blocks as the temporal convolutional network, and the DCN symmetrically fills both sides of the feature map.

5. The method for deinterleaving deep learning radar signals based on differential attention assistance according to claim 4, characterized in that, After deep feature extraction from multi-scale sequences, the process also includes reconstructing the time series: The time-series modeling module is used to capture long-term and short-term dependencies in signal sequences. Position encoding is introduced into the time-series modeling module to provide position information for each time step in the sequence. Then, a bidirectional long short-term memory network is used to capture forward and backward dependencies. Position encoding is implemented using sine and cosine functions.

6. The method for deinterleaving deep learning radar signals based on differential attention assistance according to claim 5, characterized in that, For the differential attention mechanism, the specific implementation is as follows: The arrival time is differentially analyzed to obtain the PRI parameters. By determining the PRI modulation scheme, signals from different transmitters are separated from the interleaved signals. The differential attention mechanism combines standard attention calculation with differential feature modeling. Standard attention calculates the similarity between features through query-key-value projection. The calculation process of the standard attention score is shown in the following formula: in, It is the input feature representation. These are the linear projection parameters of the query and the key. It is a query and key representation. It is an attention score matrix. It is a standardization factor to ensure gradient stability; Differential attention explicitly computes temporal difference features to capture similarity patterns in pulse intervals. The difference features are obtained by subtracting features from those of adjacent time steps. After feature encoding, a similarity matrix is ​​calculated. The calculation process is shown below: It is the difference vector between adjacent time steps. It is the parameter matrix of the differential feature encoder. Representation layer normalization operations ensure training stability. It is the SiLU activation function, where It is the sigmoid function. These are the encoded differential features. It is a difference similarity matrix. It is a normalization factor that ensures the stability of the dot product value; Finally, the standard attention score and the difference similarity matrix are weighted and fused, and then normalized by Softmax for use in weighted aggregate features.

7. The method for deinterleaving deep learning radar signals based on differential attention assistance according to claim 6, characterized in that, For the dataset obtained in S2, including the training set and the validation set, the deep neural network model is trained using the training set and evaluated using the validation set. The model parameters are continuously adjusted to obtain the optimal neural network model, which is then used to de-interleave the aliased DTOA sequences.

Images