Two-channel deep learning sorting method for intra-pulse and inter-pulse feature fusion

Abstract

The invention discloses a dual-channel deep learning sorting method for intra-pulse and inter-pulse feature fusion, and the method changes a conventional serial architecture, and constructs an end-to-end parallel processing frame based on feature fusion. And a cross-modal attention mechanism is innovatively introduced, so that deep interaction and collaborative reasoning of intra-pulse modulation features and inter-pulse statistical features are realized. According to the method, the limitation of a single feature dimension is effectively overcome, and the sorting accuracy and robustness in a parameter fuzzy scene are improved.

Description

Technical Field

[0001] This invention belongs to the field of radar technology, specifically relating to a dual-channel deep learning sorting method that fuses intra-pulse and inter-pulse features. Background Technology

[0002] Radar signal sorting is a crucial component of electronic reconnaissance systems, serving as an important prerequisite and foundation for radar signal feature extraction and target identification. With the increasing complexity of the modern electromagnetic environment, signals are characterized by high density, high agility, and diverse modulation modes, posing significant challenges to signal sorting technology.

[0003] Existing radar signal sorting technologies can be mainly divided into the following three categories: The first category is the traditional method based on inter-pulse parameter statistics. This type of method mainly relies on single-dimensional features such as the pulse repetition interval (PRI) for histogram statistics or sequence search. However, when faced with scenarios where modern radar signal parameters overlap significantly, the limitations of relying on only a single feature dimension become apparent, resulting in a significant decrease in sorting accuracy and a tendency to produce "missed batches" and "incorrect batches".

[0004] The second category is end-to-end sorting methods based on deep learning. These methods utilize deep neural networks to directly extract features from the pulse stream and learn sorting rules. For example, they use recurrent neural networks (RNNs) to perform time-of-arrival (TOA) sequence modeling, or map the pulse stream data into image or point cloud representations and then use segmentation networks for clustering. While this type of method effectively reduces the reliance on manual feature extraction, its core mechanism is unsupervised or self-supervised clustering, which can only achieve blind separation of pulse sequences and lacks semantic understanding of key attributes such as modulation type and PRI mode. The system still requires secondary processing by a backend recognition module, increasing the complexity and latency of the processing chain, making it difficult to meet the integrated real-time processing requirements in complex electromagnetic environments.

[0005] The third type is the cascaded method that combines inter-pulse and intra-pulse features. This method employs a serial processing architecture, first using inter-pulse features for deinterlacing, and then using intra-pulse features for modulation and recognition. Theoretically, it has a higher recognition ceiling, but its cascaded structure introduces the problem of "error accumulation." When the performance of the front-end inter-pulse deinterlacing stage degrades due to parameter ambiguity, it limits the overall accuracy. Furthermore, due to the lack of coordination and information feedback between inter-pulse and intra-pulse features, there is a significant bottleneck in sorting performance when dealing with complex overlapping signals.

[0006] In summary, existing technologies have technical shortcomings in handling parameter ambiguity, single feature dimensions, and structural error accumulation. There is an urgent need for a new sorting method that can synergistically utilize multi-dimensional information and avoid error accumulation. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this invention provides a dual-channel deep learning sorting method that fuses intra-pulse and inter-pulse features. This method changes the traditional serial architecture and constructs an end-to-end parallel processing framework based on feature fusion. Furthermore, it innovatively introduces a cross-modal attention mechanism, realizing deep interaction and collaborative reasoning between intra-pulse modulation features and inter-pulse statistical features. This invention effectively overcomes the limitations of a single feature dimension and improves sorting accuracy and robustness in scenarios with ambiguous parameters.

[0008] The technical solution adopted by this invention to solve its technical problem is as follows: Step 1: Parallel extraction and preprocessing of heterogeneous feature maps; For the received radar pulse sequence, two heterogeneous feature maps, intra-pulse and inter-pulse, are extracted in parallel and converted into a two-dimensional image format suitable for convolutional neural network processing; Step 1-1: Intra-pulse feature map extraction; Perform continuous wavelet transform on the baseband waveform signal of each pulse to generate a two-dimensional time-frequency map that can characterize its intra-pulse modulation information; Step 1-2: Inter-pulse feature map extraction; extract the TOA of the pulse sequence, and use the PRI transform algorithm to convert the one-dimensional TOA sequence into a two-dimensional time-PRI map that can characterize its PRI modulation law; Steps 1-3: Feature map preprocessing; The generated intra-pulse feature maps and inter-pulse feature maps are preprocessed in a targeted manner, including logarithmic transformation, edge cropping, filtering and denoising, size normalization and grayscale enhancement operations, to obtain network input with a uniform format. Step 2: Dual-channel deep feature learning; The preprocessed intra-pulse feature maps and inter-pulse feature maps are input into a parallel dual-channel deep convolutional neural network architecture to independently extract deep abstract features. Step 2-1: Construct a parallel backbone network; use two parallel deep convolutional neural networks as feature extractors, receiving intra-pulse feature maps and inter-pulse feature maps as inputs respectively; Step 2-2: Deep feature vector generation; Each backbone network performs deep learning of the patterns in the input image through multi-layer convolution and pooling operations, and finally outputs high-dimensional intra-pulse modulation deep feature vector and inter-pulse temporal regularity deep feature vector, which serve as the basis for subsequent cross-modal fusion.

[0009] Step 3: Cross-modal attention feature fusion and enhancement; The cross-modal attention fusion module is used to perform asymmetric bidirectional interaction on the intra-pulse depth features and inter-pulse depth features extracted by dual channels, so as to achieve deep complementarity and mutual enhancement of information. Step 3-1: Construct a bidirectional cross-attention mechanism; The bidirectional cross-attention mechanism achieves feature enhancement through asymmetric interaction: First, intra-pulse features actively query information from inter-pulse features, and calculate the intra-pulse feature representation that is effectively refined and enhanced by the inter-pulse information based on the correlation information of the inter-pulse features; then, inter-pulse features actively query information from intra-pulse features, and calculate the inter-pulse feature representation that is effectively refined and enhanced by the intra-pulse information. Step 3-2: Fuse and enhance features; Through bidirectional interaction, two mutually reinforcing feature vectors are finally obtained, namely, enhanced intra-pulse features and enhanced inter-pulse features. The two are then fused to obtain a final feature representation that includes comprehensive information from both intra-pulse and inter-pulse aspects. Step 4: Multi-task parallel decoding and end-to-end output of fused features; The fused feature vector is fed into two independent classification heads and decoded in parallel using a multi-task learning approach, while simultaneously outputting the pulse sorting and recognition results. Step 4-1: Parallel classifier design; Two independent classification modules are constructed, serving as intra-pulse modulation recognition head and inter-pulse modulation recognition head, respectively. Step 4-2: Output the sorting results; Ultimately, for each input pulse, the model can simultaneously provide its intra-pulse modulation type and PRI modulation type, thus completing the signal sorting task; Step 4-3: Joint optimization training; By jointly optimizing the loss functions of the two classification tasks, the entire network is trained within a unified end-to-end framework.

[0010] Preferably, step 1 specifically comprises: For the received radar pulse sequence Time-frequency diagrams are generated using continuous wavelet transform. Specifically, it is expressed as follows:

[0011] in, It is a wavelet function The complex conjugate function, wavelet function Mother wavelet After scale Translation parameters The transformation yields the following specific representation:

[0012] The generated CWT amplitude matrix is ​​the intrapulse feature map. ; Pulse arrival time series Modeled as a one-dimensional pulse sequence signal :

[0013] On a fine time grid Sampling is performed to obtain discrete time series. :

[0014] Among them, discrete sampling time points With arrival time series The relationship is:

[0015] in, For time resolution; right The time-frequency energy distribution matrix is ​​obtained by performing a short-time Fourier transform. :

[0016] in, It's frequency. It is a window function. This indicates the number of samples that are slid forward each time the window moves. It is the length of the window function; frequency axis Converting to the PRI axis yields a time-PRI two-dimensional image, which is the interpulse feature map. ; Both feature maps underwent feature preprocessing. For the intrapulse feature map, logarithmic transformation and clipping of both ends of the time axis were performed. For the interpulse feature map, a median filter was used for smoothing, and morphological closing operations were applied to fill in the small discontinuities in the spectral lines caused by pulse loss, thus enhancing the structure. Finally, both feature maps were normalized to 224 using the LANCZOS interpolation algorithm. 224 pixels, converted to 8-bit grayscale, and histogram equalization is applied to maximize image contrast.

[0017] Preferably, step 2 specifically comprises: The backbone network uses a ResNet-18 with the FC layer removed; input and Each feature vector is passed through its respective ResNet-18 channel and then subjected to adaptive global average pooling to output a 512-dimensional deep feature vector. and .

[0018] Preferably, in step 3, a cross-attention mechanism based on scaled dot product attention is used, utilizing... Enhance The calculation process is as follows: Input feature vector and Mapped to query vectors respectively Key vector Sum value vector Three semantic spaces:

[0019]

[0020]

[0021] in, , and It is a learnable weight matrix; Calculate and normalize attention weights; calculate and transpose The dot product of the query and each key yields the similarity between the query and each key, i.e., the attention score. This score matrix quantifies the similarity between each part and the query. The degree of correlation between each part. Divide the attention score by a scaling factor. To prevent the gradient from becoming too small, it is then normalized to the attention weight matrix using the Softmax function. :

[0022] in, It is the dimension of the key vector. =256; Using this weight matrix value vector Weighted summation is performed to obtain the final enhanced feature output. :

[0023] Through symmetric calculation, i.e. As , As and Interpulse enhancement features were obtained. Ultimately merged and get .

[0024] Preferably, step 4 specifically comprises: The intra-pulse modulation classification head adopts a two-layer MLP structure; the inter-pulse PRI classification head adopts a three-layer MLP structure, integrating batch normalization and Dropout layers. End-to-end training is achieved by jointly optimizing the loss functions of the two tasks. The loss function is defined as follows:

[0025] in, and These represent the true labels of the intra-pulse modulation task and the predicted logits of the network output, respectively. and The true labels and predicted logits corresponding to the inter-pulse PRI task; This is the weighting coefficient used to balance the two tasks, set to 1.0; The standard cross-entropy loss function is calculated as follows:

[0026] in, The number of training samples, It is the model on the sample The prediction results It is a sample The true label; For each input pulse, the model can simultaneously output its intra-pulse modulation type and PRI modulation type, thus completing the radar signal sorting task.

[0027] An electronic device includes: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to enable the electronic device to perform the above-described dual-channel deep learning sorting method.

[0028] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described dual-channel deep learning sorting method.

[0029] A chip includes a processor for retrieving and running a computer program from a memory, causing a device equipped with the chip to perform the dual-channel deep learning sorting method described above.

[0030] A computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the aforementioned dual-channel deep learning sorting method.

[0031] The beneficial effects of this invention are as follows: 1. Innovative parallel processing framework to avoid error accumulation: This invention adopts an end-to-end parallel processing framework, which solves the error accumulation problem in traditional cascade sorting methods and significantly improves the overall sorting accuracy and performance ceiling of the system.

[0032] 2. Deep fusion of multi-dimensional features to improve sorting performance in complex scenarios: By deeply fusing two heterogeneous information, intra-pulse waveform details and inter-pulse temporal patterns, and using cross-modal attention mechanism to achieve dynamic feature enhancement, the problem of insufficient sorting ability in scenarios with highly overlapping and unclear parameters by relying solely on a single PRI feature is effectively solved, and the sorting accuracy in complex electromagnetic environments is greatly improved.

[0033] 3. Strong robustness in sorting and efficient integrated sorting: Under harsh conditions such as low signal-to-noise ratio, the model can effectively focus on key discriminative features, suppress noise interference, and enhance anti-interference and generalization capabilities. Through end-to-end parallel processing, highly efficient integrated sorting and recognition is achieved.

[0034] This invention achieves end-to-end sorting of mixed pulse signals, optimizes the processing flow, and improves sorting efficiency. Experiments show that the method achieves a sorting accuracy of 95.4% on diverse and complex datasets, which is 9.8% higher than cascaded deep learning models and 1.4% higher than Cross-modal Transformer models. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the feature map preprocessing process proposed in this invention; Figure 2 This is a diagram of the JDR-Net network model framework proposed in this invention; Figure 3 This is a schematic diagram of the internal structure of the scaled dot product attention unit used in this invention; Figure 4 This is a comparison curve of the inter-pulse recognition accuracy of the method of this invention (with / without attention mechanism) and the cascaded model under different signal-to-noise ratios; Figure 5 shows a comparison of the inter-pulse recognition confusion matrix at a signal-to-noise ratio of 5dB. Figure 5(a) shows the result without the attention mechanism, and Figure 5(b) shows the result with the attention mechanism. Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0037] To address the performance degradation of existing radar signal sorting methods in complex electromagnetic environments due to high parameter overlap, and the error accumulation effect of traditional cascaded processing architectures, this invention proposes a dual-channel deep learning sorting method that fuses intra-pulse and inter-pulse features. This method departs from the traditional serial architecture, constructing an end-to-end parallel processing framework based on feature fusion; and innovatively introduces a cross-modal attention mechanism, enabling deep interaction and collaborative reasoning between intra-pulse modulation features and inter-pulse statistical features. This invention effectively overcomes the limitations of a single feature dimension, improving sorting accuracy and robustness in scenarios with ambiguous parameters.

[0038] To achieve the above objectives, the technical solution adopted by the present invention mainly includes the following steps: Step 1: Parallel Extraction and Preprocessing of Heterogeneous Features. For the received radar pulse sequence, both intra-pulse and inter-pulse heterogeneous features are extracted in parallel and converted into a two-dimensional image format suitable for convolutional neural network processing. The specific steps are as follows: Step 1-1: Intra-pulse feature map extraction. Perform continuous wavelet transform on the baseband waveform signal of each pulse to generate a two-dimensional time-frequency map that can accurately characterize its intra-pulse modulation information.

[0039] Step 1-2: Inter-pulse feature map extraction. Extract the TOA of the pulse sequence and use the PRI transform algorithm to convert the one-dimensional TOA sequence into a two-dimensional time-PRI map that can characterize its PRI modulation pattern.

[0040] Steps 1-3: Feature Map Preprocessing. The generated intra-pulse and inter-pulse feature maps undergo targeted preprocessing, including logarithmic transformation, edge cropping, filtering and denoising, size normalization, and grayscale enhancement, to obtain high-quality, uniformly formatted network input.

[0041] Step 2: Dual-channel deep feature learning. The two preprocessed feature maps are input into a parallel dual-channel deep convolutional neural network architecture to independently extract deep abstract features. The specific implementation steps are as follows: Step 2-1: Construct a parallel backbone network. Two parallel deep convolutional neural networks are used as feature extractors, receiving intra-pulse and inter-pulse feature maps as input, respectively.

[0042] Step 2-2: Deep Feature Vector Generation. Each backbone network performs deep learning on the patterns in the input image through multiple convolution and pooling operations, and finally outputs high-dimensional intra-pulse modulation deep feature vectors and inter-pulse temporal pattern deep feature vectors, which serve as the basis for subsequent cross-modal fusion.

[0043] Step 3: Cross-modal attention feature fusion and enhancement. Using the designed cross-modal attention fusion module, asymmetric bidirectional interaction is performed on the intra-pulse depth features and inter-pulse depth features extracted from the dual channels to achieve deep information complementarity and mutual enhancement. The specific implementation steps are as follows: Step 3-1: Construct a bidirectional cross-attention mechanism. This mechanism achieves feature enhancement through asymmetric interaction: First, intra-pulse features actively query information from inter-pulse features, and calculate the intra-pulse feature representation effectively refined and enhanced by the inter-pulse information based on the correlation information of the inter-pulse features; subsequently, inter-pulse features actively query information from intra-pulse features, and calculate the inter-pulse feature representation effectively refined and enhanced by the intra-pulse information.

[0044] Step 3-2: Feature Fusion and Enhancement. Through the bidirectional interaction described above, the model ultimately obtains two mutually enhancing feature vectors: enhanced intra-pulse features and enhanced inter-pulse features. These two vectors are then fused to obtain a final feature representation that incorporates both intra-pulse and inter-pulse information.

[0045] Step 4: Multi-task parallel decoding and end-to-end output of fused features. The fused feature vector is fed into two independent classification heads for parallel decoding using a multi-task learning approach, while simultaneously outputting the pulse sorting and recognition results.

[0046] Step 4-1: Parallel Classifier Design. Construct two independent classification modules, serving as the intra-pulse modulation recognition head and the inter-pulse modulation recognition head, respectively.

[0047] Step 4-2: Outputting the sorting results. Finally, for each input pulse, the model can simultaneously output its intra-pulse modulation type and PRI modulation type, thus completing the signal sorting task.

[0048] Step 4-3: Joint Optimization Training. The loss functions for the two classification tasks are jointly optimized, training the entire network within a unified end-to-end framework. This multi-task learning strategy forces the network to learn shared feature representations applicable to both tasks, improving the model's overall generalization ability and discrimination accuracy.

[0049] Example: To verify the effectiveness of the proposed method, this example simulates and generates a large-scale, diverse radar pulse dataset. The dataset includes all combinations of three typical intra-pulse modulation types (LFM, BPSK, CW) and five PRI modulation types (fixed PRI, jittery PRI, slip PRI, staggered PRI, and agile PRI). Signal parameters such as pulse width, bandwidth, and PRI range are completely randomized within preset ranges. The signal-to-noise ratio is distributed between 0 dB and 20 dB, and pulse loss is introduced to simulate a real, complex electromagnetic environment. Randomly generated sorting scenarios with overlapping signals from different radiation sources are used, with each simulation scenario containing 2–5 radars. Detailed parameters for each signal source are randomly generated within preset ranges; specific parameter settings are shown in Table 1.

[0050] Table 1. Simulation inter-pulse characteristic parameter settings for radar signals

[0051] The dataset generated 30,000 valid pulse samples, which were divided into training, validation, and test sets according to a ratio of 70%, 15%, and 15%, respectively.

[0052] The present invention proposes A dual-channel deep learning sorting method that fuses intra- and inter-pulse features The process includes feature extraction and the construction and training of the network model. The specific implementation steps are as follows: Step 1: For the received radar pulse sequence, extract both intra-pulse and inter-pulse heterogeneous features in parallel, and convert them into a two-dimensional image format suitable for convolutional neural network processing.

[0053] Step 1-1: Process the received radar pulse sequence Time-frequency diagrams are generated using continuous wavelet transform. Specifically, it is expressed as follows:

[0054] in, It is a wavelet function The complex conjugate function, wavelet function Mother wavelet After scale Translation parameters The transformation yields the following specific representation:

[0055] The generated CWT amplitude matrix is ​​the intrapulse feature map. .

[0056] Steps 1-2: First, sequence the pulse arrival times. Modeled as a one-dimensional pulse sequence signal :

[0057] On a fine time grid Sampling is performed to obtain discrete time series. :

[0058] Among them, discrete sampling time points With arrival time series The relationship is:

[0059] in, For time resolution.

[0060] right The time-frequency energy distribution matrix is ​​obtained by performing a short-time Fourier transform. :

[0061] in, It's frequency. It is a window function. This indicates the number of samples that are slid forward each time the window moves. It is the length of the window function.

[0062] Finally, the frequency axis Converting to the PRI axis yields a time-PRI two-dimensional image, which is the interpulse feature map. .

[0063] Steps 1-3: Both feature maps are processed Figure 1 The preprocessing flow shown performs feature preprocessing. For intrapulse feature maps, logarithmic transformation and clipping of both ends of the time axis are performed. For interpulse feature maps, a median filter is used for smoothing, and morphological closing operations are applied to fill in the small discontinuities in the spectral lines caused by pulse loss, thus enhancing the structure. Finally, both feature maps are normalized to 224 using the LANCZOS interpolation algorithm. 224 pixels, converted to 8-bit grayscale, and histogram equalization is applied to maximize image contrast.

[0064] Step 2: Construct a parallel dual-channel convolutional neural network to perform independent deep feature learning on the two preprocessed feature maps.

[0065] Step 2-1: The backbone network uses ResNet-18 (FC layer removed). Input and Each feature vector is passed through its respective ResNet-18 channel and then subjected to adaptive global average pooling to output a 512-dimensional deep feature vector. and .

[0066] Step 3: This module employs a cross-attention mechanism based on Scaled Dot-Product Attention to utilize... Enhance For example, the calculation process is as follows: Step 3-1: Input feature vector and Mapped to query vectors respectively Key vector Sum value vector Three semantic spaces ( =256):

[0067]

[0068]

[0069] in, , and It is a learnable weight matrix.

[0070] Step 3-2: Calculate and normalize the attention weights. and transpose The dot product of the query and each key yields the similarity between the query and each key, i.e., the attention score. This score matrix quantifies the similarity between each part and the query. The degree of correlation between each part. Divide the attention score by a scaling factor. To prevent the gradient from becoming too small, it is then normalized to the attention weight matrix using the Softmax function. :

[0071] in, It is the dimension of the key vector.

[0072] Using this weight matrix value vector Weighted summation is performed to obtain the final enhanced feature output. :

[0073] Step 3-3: Through symmetric calculation, i.e. As , As and Interpulse enhancement features were obtained. Ultimately merged and get .

[0074] Step 4: Multi-task parallel decoding and output. The fused feature vectors... The pulses are fed into two independent classification heads, one for intrapulse and one for interpulse, and decoded in parallel using a multi-task learning approach, while simultaneously outputting the pulse sorting and recognition results.

[0075] Step 4-1: The intra-pulse modulation classification head adopts a simple two-layer MLP structure; the inter-pulse PRI classification head adopts a three-layer MLP structure, integrating batch normalization (BatchNorm) and dropout layers to enhance the model's ability to model complex PRI patterns and prevent overfitting.

[0076] Step 4-2: Achieve end-to-end training by jointly optimizing the loss functions of the two tasks. The loss function is defined as follows:

[0077] in, and These represent the true labels of the intra-pulse modulation task and the predicted logits of the network output, respectively. and This corresponds to the true label and predicted logits of the inter-pulse PRI task. These are the weighting coefficients used to balance the two tasks, and in this study, they are all set to 1.0; The standard cross-entropy loss function is calculated as follows:

[0078] in, The number of training samples, It is the model on the sample The prediction results It is a sample The true label.

[0079] Finally, for each input pulse, the model can simultaneously output its intra-pulse modulation type and PRI modulation type, thus completing the radar signal sorting task.

[0080] The overall performance was evaluated on a general test set and compared with the Cross-modal Transformer model. This model uses the same dual-channel ResNet-18 backbone for feature extraction as the one described in this invention, and the fusion module is a standard Transformer Encoder. It treats intra- and inter-pulse feature vectors as sequences and uses a self-attention mechanism for interactive fusion, representing a mainstream feature-level fusion technique. To ensure fairness in the comparison, all models were trained and tested on the same dataset constructed in this invention, and the results are shown in Table 2.

[0081] Table 2 Performance Comparison of Multiple Models

[0082] The comprehensive performance comparison shows that the parallel fusion framework based on cross-modal attention proposed in this invention has significant advantages. Compared with the cascaded sorting model, the sorting accuracy of the method in this invention is improved by 9.8%, and compared with the Cross-modal Transformer model, the sorting accuracy of the method in this invention is improved by 1.4%. This verifies the effectiveness of the method in structural design and fusion mechanism in this invention. The specific analysis is as follows: (1) The parallel fusion framework is superior to the cascaded structure. It fundamentally avoids the error accumulation problem inherent in the cascaded process and achieves a higher performance ceiling; (2) The cross-modal attention mechanism is superior to the self-attention mechanism. The self-attention mechanism of the Cross-modal Transformer model needs to learn complex dependencies within the sequence containing all information. In contrast, the cross-attention mechanism adopted by the cross-modal attention mechanism in this invention has a clearer goal: to use one modality to query and enhance another modality. For short sequence tasks containing only two heterogeneous vectors, this guided interaction is more efficient and targeted, and can achieve accurate cross-domain information enhancement at a lower cost.

[0083] To evaluate the model's performance at different signal-to-noise ratios (SNRs) and measure its robustness in real-world environments, the test set was divided into several intervals based on SNR. The inter-pulse recognition accuracy of each model in each interval was calculated, and the results are as follows: Figure 4 As shown. From Figure 4The performance curves show that in the low signal-to-noise ratio (SNR) region (0–10 dB), the accuracy of the proposed method is higher than that of the cascaded model, regardless of whether an attention mechanism is introduced. Specifically, the model with the attention mechanism not only significantly outperforms the cascaded model but also the Cross-modal Transformer model, and this gap increases as the SNR decreases. This indicates that under conditions of feature ambiguity and low SNR, cross-modal attention can more intelligently perform information filtering and fusion, improving system robustness. As the SNR increases, when the signal clarity is high (>15 dB), the accuracy of all models tends to converge, but the accuracy of the proposed method consistently remains the highest, verifying its comprehensive advantages under various conditions.

[0084] To further visually verify the role of cross-modal attention mechanisms under low signal-to-noise ratio (SNR) conditions, the inter-pulse recognition confusion matrices of JDR-Net models with and without attention mechanisms were compared on the SNR=5dB dataset, as shown in Figure 5. Figure 5(a) shows the model without attention mechanisms, and Figure 5(b) shows the model with attention mechanisms. The confusion matrix results show that the model with attention mechanisms achieves higher correct recognition rates for slip, staggered, and group variation types than the model without attention mechanisms, while fixed and jitter types remain unchanged or show slight improvement. These results indicate that attention mechanisms help the model more accurately capture and distinguish subtle differences in complex PRI patterns under low SNR conditions, demonstrating stronger robustness by intelligently weighting and amplifying effective features while suppressing noise interference.

Claims

1. A dual-channel deep learning sorting method that fuses intra- and inter-pulse features, characterized in that, Includes the following steps: Step 1: Parallel extraction and preprocessing of heterogeneous feature maps; For the received radar pulse sequence, two heterogeneous feature maps, intra-pulse and inter-pulse, are extracted in parallel and converted into a two-dimensional image format suitable for convolutional neural network processing; Step 1-1: Intra-pulse feature map extraction; Perform continuous wavelet transform on the baseband waveform signal of each pulse to generate a two-dimensional time-frequency map that can characterize its intra-pulse modulation information; Step 1-2: Inter-pulse feature map extraction; extract the TOA of the pulse sequence, and use the PRI transform algorithm to convert the one-dimensional TOA sequence into a two-dimensional time-PRI map that can characterize its PRI modulation law; Steps 1-3: Feature map preprocessing; The generated intra-pulse feature maps and inter-pulse feature maps are preprocessed in a targeted manner, including logarithmic transformation, edge cropping, filtering and denoising, size normalization and grayscale enhancement operations, to obtain network input with a uniform format. Step 2: Dual-channel deep feature learning; The preprocessed intra-pulse feature maps and inter-pulse feature maps are input into a parallel dual-channel deep convolutional neural network architecture to independently extract deep abstract features. Step 2-1: Construct a parallel backbone network; use two parallel deep convolutional neural networks as feature extractors, receiving intra-pulse feature maps and inter-pulse feature maps as inputs respectively; Step 2-2: Deep feature vector generation; Each backbone network performs deep learning of the patterns in the input image through multi-layer convolution and pooling operations, and finally outputs high-dimensional intra-pulse modulation deep feature vector and inter-pulse temporal regularity deep feature vector, which serve as the basis for subsequent cross-modal fusion. Step 3: Cross-modal attention feature fusion and enhancement; The cross-modal attention fusion module is used to perform asymmetric bidirectional interaction on the intra-pulse depth features and inter-pulse depth features extracted by dual channels, so as to achieve deep complementarity and mutual enhancement of information. Step 3-1: Construct a bidirectional cross-attention mechanism; The bidirectional cross-attention mechanism achieves feature enhancement through asymmetric interaction: First, intra-pulse features actively query information from inter-pulse features, and calculate the intra-pulse feature representation that is effectively refined and enhanced by inter-pulse information based on the correlation information of inter-pulse features. Subsequently, the interpulse features actively query the intrapulse features for information, and calculate the interpulse feature representation that is effectively refined and enhanced by the intrapulse information; Step 3-2: Fuse and enhance features; Through bidirectional interaction, two mutually reinforcing feature vectors are finally obtained, namely, enhanced intra-pulse features and enhanced inter-pulse features. The two are then fused to obtain a final feature representation that includes comprehensive information from both intra-pulse and inter-pulse aspects. Step 4: Multi-task parallel decoding and end-to-end output of fused features; The fused feature vector is fed into two independent classification heads and decoded in parallel using a multi-task learning approach, while simultaneously outputting the pulse sorting and recognition results. Step 4-1: Parallel classifier design; Two independent classification modules are constructed, serving as intra-pulse modulation recognition head and inter-pulse modulation recognition head, respectively. Step 4-2: Output the sorting results; Ultimately, for each input pulse, the model can simultaneously provide its intra-pulse modulation type and PRI modulation type, thus completing the signal sorting task; Step 4-3: Joint optimization training; By jointly optimizing the loss functions of the two classification tasks, the entire network is trained within a unified end-to-end framework.

2. The dual-channel deep learning sorting method based on the fusion of intra-pulse and inter-pulse features according to claim 1, characterized in that, Step 1 specifically involves: For the received radar pulse sequence Time-frequency diagrams are generated using continuous wavelet transform. Specifically, it is expressed as follows: in, It is a wavelet function The complex conjugate function, wavelet function Mother wavelet After scale Translation parameters The transformation yields the following specific representation: The generated CWT amplitude matrix is ​​the intrapulse feature map. ; Pulse arrival time series Modeled as a one-dimensional pulse sequence signal : On a fine time grid Sampling is performed to obtain discrete time series. : Among them, discrete sampling time points With arrival time series The relationship is: in, For time resolution; right The time-frequency energy distribution matrix is ​​obtained by performing a short-time Fourier transform. : in, It's frequency. It is a window function. This indicates the number of samples that are slid forward each time the window moves. It is the length of the window function; frequency axis Converting to the PRI axis yields a time-PRI two-dimensional image, which is the interpulse feature map. ; Both feature maps underwent feature preprocessing. For the intrapulse feature map, logarithmic transformation and clipping of both ends of the time axis were performed. For the interpulse feature map, a median filter was used for smoothing, and morphological closing operations were applied to fill in the small discontinuities in the spectral lines caused by pulse loss, thus enhancing the structure. Finally, both feature maps were normalized to 224 using the LANCZOS interpolation algorithm. 224 pixels, converted to 8-bit grayscale, and histogram equalization is applied to maximize image contrast.

3. The dual-channel deep learning sorting method based on the fusion of intra- and inter-pulse features according to claim 2, characterized in that, Step 2 specifically involves: The backbone network uses a ResNet-18 with the FC layer removed; input and Each feature vector is passed through its respective ResNet-18 channel and then subjected to adaptive global average pooling to output a 512-dimensional deep feature vector. and .

4. The dual-channel deep learning sorting method for fusing intra- and inter-pulse features according to claim 3, characterized in that, In step 3, a cross-attention mechanism based on scaled dot product attention is used, utilizing... Enhance The calculation process is as follows: Input feature vector and Mapped to query vectors respectively Key vector Sum value vector Three semantic spaces: in, , and It is a learnable weight matrix; Calculate and normalize attention weights; calculate and transpose The dot product of the query and each key yields the similarity between the query and each key, i.e., the attention score. This score matrix quantifies the similarity between each part and the query. The degree of correlation between each part; dividing the attention score by a scaling factor. To prevent the gradient from becoming too small, it is then normalized to the attention weight matrix using the Softmax function. : in, It is the dimension of the key vector. =256; Using this weight matrix value vector Weighted summation is performed to obtain the final enhanced feature output. : Through symmetric calculation, i.e. As , As and Interpulse enhancement features were obtained. Ultimately merged and get .

5. The dual-channel deep learning sorting method based on the fusion of intra-pulse and inter-pulse features according to claim 4, characterized in that, Step 4 specifically involves: The intra-pulse modulation classification head adopts a two-layer MLP structure; the inter-pulse PRI classification head adopts a three-layer MLP structure, integrating batch normalization and Dropout layers. End-to-end training is achieved by jointly optimizing the loss functions of the two tasks. The loss function is defined as follows: in, and These represent the true labels of the intra-pulse modulation task and the predicted logits of the network output, respectively. and The true labels and predicted logits corresponding to the inter-pulse PRI task; This is the weighting coefficient used to balance the two tasks, set to 1.0; The standard cross-entropy loss function is calculated as follows: in, The number of training samples, It is the model on the sample The prediction results It is a sample The true label; For each input pulse, the model can simultaneously output its intra-pulse modulation type and PRI modulation type, thus completing the radar signal sorting task.

6. An electronic device, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 5.

8. A chip, characterized in that, include: A processor for retrieving and running a computer program from memory, causing a device on which the chip is mounted to perform the method as described in any one of claims 1 to 5.

9. A computer program product, characterized in that, The computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the method as described in any one of claims 1 to 5.

Images