Radar signal intra-pulse modulation recognition method based on branch feature extraction and fusion

Abstract

The application discloses a radar signal intra-pulse modulation recognition method based on branch feature extraction and fusion, comprising the following steps: acquiring a radar signal, extracting time sequence features and time-frequency features of the radar signal; performing difference operation based on the time sequence features and the time-frequency features to obtain difference features; performing point multiplication operation based on the time sequence features and the time-frequency features to obtain correlation features; splicing the time sequence features, the time-frequency features, the difference features and the correlation features to obtain an interactive enhanced feature vector; inputting the interactive enhanced feature vector into a dynamic weight generation network to obtain branch fusion weights of the interactive enhanced feature vector; performing weighted summation on the time sequence features and the time-frequency features respectively by using the branch fusion weights, and performing residual mapping on the weighted summation result to obtain final fusion features; and inputting the final fusion features into a classification network to output an intra-pulse modulation category corresponding to the radar signal. The application can improve the recognition accuracy of the radar signal.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing, specifically relating to a radar signal intra-pulse modulation identification method based on branch feature extraction and fusion. Background Technology

[0002] Intra-pulse modulation identification of radar signals is a key aspect of electronic reconnaissance. As modern radar systems increasingly adopt low probability of intercept waveform designs, the structural characteristics of radar signals in the time, frequency, and phase domains become more complex. Furthermore, in actual receiving scenarios, they are often affected by factors such as noise interference, random parameter variations, and insufficient prior information, which significantly increases the difficulty of intra-pulse modulation identification.

[0003] Currently, radar signal intra-pulse modulation identification methods mainly rely on deep learning-based automatic feature learning methods. These typically use time-series or time-frequency features as input and employ models such as convolutional neural networks and recurrent neural networks for identification.

[0004] However, most existing methods are prone to feature space distribution mismatch during identification, which leads to low accuracy in radar signal identification. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a radar signal intra-pulse modulation identification method based on branch feature extraction and fusion. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a radar signal intra-pulse modulation identification method based on branch feature extraction and fusion, comprising: acquiring radar signals, and extracting the temporal and time-frequency features of the radar signals; Difference operations are performed on the time-series features and the time-frequency features to obtain difference features; dot product operations are performed on the time-series features and the time-frequency features to obtain correlation features; the time-series features, the time-frequency features, the difference features, and the correlation features are concatenated to obtain an interactive enhancement feature vector; The interaction enhancement feature vector is input into a dynamic weight generation network to obtain the branch fusion weights of the interaction enhancement feature vector; the branch fusion weights are used to perform weighted summation on the time-series features and the time-frequency features respectively, and residual mapping is performed on the weighted summation results to obtain the final fused features; The final fused features are input into the classification network, which outputs the intra-pulse modulation category corresponding to the radar signal.

[0006] In one embodiment of the present invention, acquiring radar signals and extracting time-series and time-frequency features aligned with the scale and distribution of the radar signals includes: Acquire radar signals and convert them into complex in-phase orthogonal sequences; The complex in-phase orthogonal sequence is input into the time series feature extraction module to obtain the time series feature vector; The complex in-phase orthogonal sequence is input into the time-frequency feature extraction module to obtain the time-frequency feature vector; The temporal feature vector and the time-frequency feature vector are mapped to a unified common feature space for scale and distribution alignment to obtain the temporal and time-frequency features of the radar signal.

[0007] In one embodiment of the present invention, difference features are obtained by performing a difference operation based on the time-series features and the time-frequency features, including: Calculate the element-level absolute difference between the time-series feature and the time-frequency feature to obtain an absolute difference vector; use the absolute difference vector as the difference feature.

[0008] In one embodiment of the present invention, relevant features are obtained by performing a dot product operation based on the time-series features and the time-frequency features, including: The time-series feature and the time-frequency feature are multiplied element-wise to obtain an element-level dot product vector representing the consistency information between the time-series feature and the time-frequency feature; the element-level dot product vector is used as the relevant feature.

[0009] In one embodiment of the present invention, the time-series features, the time-frequency features, the difference features, and the correlation features are concatenated to obtain an interaction-enhanced feature vector, including: The time-series features, time-frequency features, difference features, and correlation features are concatenated and strung together element by element along the channel dimension to obtain the interaction-enhanced feature vector.

[0010] In one embodiment of the present invention, the interaction enhancement feature vector is input into a dynamic weight generation network to obtain the branch fusion weights corresponding to the interaction enhancement feature vector, including: The interaction enhancement feature vector is subjected to two-level nonlinear feature dimensionality reduction and mapping to obtain the branch score vector; The branch scoring vector is normalized to generate time-series branch weights and time-frequency branch weights, and the branch fusion weights are composed of the time-series branch weights and the time-frequency branch weights.

[0011] In one embodiment of the present invention, the interaction enhancement feature vector is subjected to two-level nonlinear feature dimensionality reduction and mapping to obtain a branch scoring vector, including: The interaction enhancement feature vector is subjected to a first-level fully connected linear mapping using the first weight matrix and the first bias term to obtain the first-level mapped feature vector; The first-level mapped feature vector is nonlinearly activated using a linear rectified activation function to output intermediate dimensionality-reduced features. The branch score vector is obtained by performing a second-level fully connected linear mapping on the intermediate dimensionality-reduced features using the second weight matrix and the second bias term.

[0012] In one embodiment of the present invention, the branch fusion weights are used to perform weighted summation on the time-series features and the time-frequency features respectively, and residual mapping is performed on the weighted summation results to obtain the final fused features, including: The element-wise product of the temporal branch weights and the temporal features is performed to obtain the first product result. The time-frequency branch weights are element-wise multiplied with the time-frequency features to obtain a second product result. The first product result and the second product result are summed to obtain the coarse fusion feature vector; The coarse fused feature vector is input into the residual refinement network, which outputs intermediate enhanced features. Linear mapping is performed on the intermediate enhancement features to obtain deep refined features; The coarse fused feature vector and the deep refined feature vector are summed element-wise to obtain the final fused feature.

[0013] In one embodiment of the present invention, the residual refining network includes: The residual refinement network consists of a linear layer, a layer normalization layer, a linear rectified activation layer, and a random deactivation layer connected in sequence; wherein the input dimension and output dimension of the residual refinement network are both 128.

[0014] In one embodiment of the present invention, the classification network includes: The system comprises a fully connected layer, a batch normalization layer, a linear rectification activation layer, a random deactivation layer, and a fully connected output layer, connected sequentially. The fully connected layer has an output dimension of 256, and the fully connected output layer has an output dimension of 15, corresponding to 15 categories of intra-pulse modulation for radar signals. Compared with existing technologies, the advantages of this invention are: This invention extracts the temporal and time-frequency features of radar signals, and then performs difference and dot product operations on these features to obtain difference and correlation features. These features are then concatenated to obtain an interactively enhanced feature vector. This avoids the problem of existing methods focusing on a single representation domain, thus improving feature utilization, reducing the impact of feature space distribution mismatch, and ultimately enhancing the accuracy of radar signal recognition.

[0015] Simultaneously, this invention inputs the interactive enhancement feature vector into a dynamic weight generation network to obtain the branch fusion weights of the interactive enhancement feature vector. The branch fusion weights are then used to perform weighted summation on the temporal features and the time-frequency features respectively, and residual mapping is applied to the weighted summation results to obtain the final fused features. This avoids the problem of existing methods simply concatenating temporal and time-frequency features, thereby solving the feature space distribution mismatch problem and improving the recognition accuracy of radar signals.

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the radar signal intra-pulse modulation identification method based on branch feature extraction and fusion provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the architecture of the radar signal intra-pulse modulation recognition system based on branch feature extraction and fusion provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the temporal feature extraction module architecture of the radar signal intra-pulse modulation identification method based on branch feature extraction and fusion provided in this embodiment of the invention; Figure 4 This is a schematic diagram of the time-frequency feature extraction module architecture of the radar signal intra-pulse modulation identification method based on branch feature extraction and fusion provided in this embodiment of the invention; Figure 5 This is a graph showing the recognition accuracy of the radar signal intra-pulse modulation recognition method based on branch feature extraction and fusion provided in this embodiment of the invention. Figure 6 This is a confusion matrix diagram of the radar signal intra-pulse modulation identification method based on branch feature extraction and fusion provided in the embodiments of the present invention; Figure 7 This is a comparison of ablation experiments of the radar signal intra-pulse modulation identification method based on branch feature extraction and fusion provided in this embodiment of the invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0019] This invention provides a method for identifying intra-pulse modulation of radar signals based on branch feature extraction and fusion. Please refer to... Figure 1 , Figure 1 This is a flowchart illustrating a radar signal intra-pulse modulation identification method based on branch feature extraction and fusion, provided by an embodiment of the present invention.

[0020] Before introducing the method's workflow, we will first describe the radar signal intra-pulse modulation recognition system based on branch feature extraction and fusion that this method utilizes. The architecture of this system can be found in [link to system architecture]. Figure 2 As shown.

[0021] The radar signal intra-pulse modulation recognition system based on branch feature extraction and fusion includes a dual-branch feature extraction module, a feature interaction enhancement module, an adaptive fusion module, and a classification module.

[0022] After receiving the radar signal, the system transmits the radar signal to the dual-branch feature extraction module. The dual-branch feature extraction module extracts the time-series and time-frequency features of the radar signal and outputs the time-series and time-frequency features to the feature interaction enhancement module.

[0023] After receiving the temporal and time-frequency features, the feature interaction enhancement module first aligns the temporal and time-frequency features in terms of scale and distribution. Then, it performs a difference operation on the temporal and time-frequency features to obtain the difference features; and performs a dot product operation on the temporal and time-frequency features to obtain the correlation features. Finally, it concatenates the temporal, time-frequency, difference, and correlation features, outputting the interactively enhanced feature vector to the adaptive fusion module.

[0024] After receiving the interactive enhancement feature vector, the adaptive fusion module first inputs the interactive enhancement feature vector into the dynamic weight generation network to obtain the branch fusion weights corresponding to the interactive enhancement feature vector. Then, it performs weighted summation on the temporal features and time-frequency features respectively, and performs residual mapping on the weighted summation result to obtain the final fused features. Finally, it outputs the final fused features to the classification module.

[0025] After receiving the final fused features, the classification module performs mapping processing on the final fused features to obtain the corresponding classification results and completes the intra-pulse modulation category determination of the radar signal based on the classification results.

[0026] It should be noted that the model used in the radar signal intra-pulse modulation recognition system based on branch feature extraction and fusion is a whole, and the whole model needs to be pre-trained before implementing the method of this invention.

[0027] During training, the temporal feature extraction module and the time-frequency feature extraction module are pre-trained separately to obtain their respective initialization parameters. Then, the pre-trained parameters are loaded into the overall model, and the overall network, including the temporal feature extraction module, the time-frequency feature extraction module, the feature interaction enhancement module, the adaptive fusion module, and the classification module, is jointly trained.

[0028] Pre-training methods include common model training techniques, which are not specifically limited here.

[0029] The following is combined Figure 1 This invention introduces a radar signal intra-pulse modulation identification method based on branch feature extraction and fusion, as provided in an embodiment of the present invention.

[0030] The method includes the following steps S1 to S4: S1. Acquire radar signals and extract time-series and time-frequency features that align with the scale and distribution of the radar signals.

[0031] Among them, temporal features are a representation of the continuous evolution trajectory of radar signals over time. Temporal features are used to characterize the intra-pulse phase transient jump law and the temporal envelope fluctuation characteristics of high-order one-dimensional vector sequences.

[0032] Time-frequency features refer to the multidimensional representation exhibited after mapping complex in-phase orthogonal sequences to a two-dimensional joint time-frequency domain through a specific joint time-frequency transform. Time-frequency features are used to characterize the frequency variation of radar signals over time and local image texture feature vectors.

[0033] The purpose of aligning the scale and distribution of time-series and time-frequency features is to eliminate the heterogeneous spatial mismatch problem caused by differences in extraction source, feature dimension, numerical unit, and data distribution between time-series and time-frequency features.

[0034] In one possible implementation, acquiring radar signals and extracting time-series and time-frequency features aligned with the scale and distribution of the radar signals includes the following steps S11 to S14: S11. Acquire radar signals and convert them into complex in-phase orthogonal sequences.

[0035] The radar signal includes a complex in-phase orthogonal sequence. This sequence comprises in-phase and quadrature components, also known as I-components and Q-components. For example, the I-components and Q-components are calculated using formula (1).

[0036] (1); in, For in-phase components, For orthogonal components, For radar signals, Let be the real part of the radar signal. This represents the imaginary part of the radar signal.

[0037] S12. Input the complex in-phase orthogonal sequence into the time series feature extraction module to obtain the time series feature vector.

[0038] Please see Figure 2 The dual-branch feature extraction module includes a temporal feature extraction module. Please refer to... Figure 3 The temporal feature extraction module includes multi-scale convolutional units, stacked recurrent units, temporal attention units, and fully connected mapping units.

[0039] Specifically, S12 includes the following steps S121 to S126: S121. Utilize multi-scale convolutional units to receive complex in-phase orthogonal sequences of radar signals, extract local temporal features under different receptive fields, and output multi-scale temporal features.

[0040] Among them, the multi-scale convolutional unit uses one-dimensional convolutional kernels with different receptive fields to scan the input sequence in parallel, and captures the intra-pulse multi-scale micro-features of radar signals at different time granularities along the time axis.

[0041] The convolutional front end of the multi-scale convolutional unit includes a multi-scale feature-aware submodule and a subsequent concatenated convolutional submodule.

[0042] The multi-scale feature perception submodule comprises three parallel convolutional branches and one pooling branch. The kernel lengths of the three parallel convolutional branches are 3, 7, and 11, respectively, with a stride of 1, and each branch has 32 output channels. The pooling branch includes a max-pooling layer and a one-dimensional convolutional layer, also with 32 output channels. The multi-scale feature perception submodule concatenates the outputs of each branch along the channel dimension to obtain 128 channels of local temporal features. These local temporal features are then fed into the subsequent concatenated convolutional submodule for processing, outputting the multi-scale temporal features.

[0043] S122. Utilize stacked cyclic units to receive multi-scale temporal features and output temporal features at different times.

[0044] Among them, the stacked loop unit is used to perform time-dimensional sequence modeling of multi-scale features through a multi-layer gated loop structure, deeply capturing the transient jump law and temporal evolution characteristics of the signal pulse phase.

[0045] Specifically, the stacked recurrent unit includes a gated recurrent unit, a long short-term memory network, and a bidirectional gated recurrent unit connected in sequence. The gated recurrent unit has a hidden layer dimension of 64, the long short-term memory network has a hidden layer dimension of 64, and the bidirectional gated recurrent unit has a hidden layer dimension of 64 in each direction.

[0046] S123. Utilize the temporal attention unit to receive temporal features at different times, and weight the temporal features at different times to obtain attention scores.

[0047] The temporal attention unit is used to adaptively calculate the association weights of features at each time point, and explicitly focuses on key time segments that are discriminative of modulation patterns through weighted aggregation, thereby suppressing redundant interference.

[0048] Specifically, the attention score is calculated according to formula (2): (2); in, For the first Attention score at each time step This is the transpose of the attention weight vector. The hyperbolic tangent activation function is used. The weight matrix is ​​a learnable matrix. For stacked cyclic units in the first The output of each time step This is a bias term.

[0049] S124. Normalize the attention score using the temporal attention unit to obtain the attention weight corresponding to each time position.

[0050] Attention weights are calculated according to formula (3): (3); in, For the first Time attention weights for each time step For the first Attention score at each time step The total length of the time step. For the first Attention score at each time step.

[0051] S125. Use the time attention unit to weight the attention weights and output a time sequence representation that highlights key time sequence information.

[0052] The timing representation is calculated according to formula (4): (4); in, For time series representation, The total length of the time step. For the first Time attention weights for each time step For stacked cyclic units in the first Output at each time step.

[0053] S126. The weighted temporal representation is mapped to a 128-dimensional temporal feature vector using a fully connected mapping unit, which is then used as the output of the temporal branch.

[0054] The fully connected mapping unit is used to perform linear matrix multiplication on the aggregated temporal features, reorganize the feature space and project it to a set dimension scale, and output the final temporal feature vector.

[0055] The temporal feature vector refers to the original time-domain specific multidimensional feature flow mined by the temporal feature extraction module, which covers intrapulse multi-scale information and transient phase evolution law, and serves as the basis for normalized multimodal features to be introduced subsequently with scale and distribution alignment mechanisms.

[0056] S13. Input the complex in-phase orthogonal sequence into the time-frequency feature extraction module to obtain the time-frequency feature vector.

[0057] Please see Figure 2 The dual-branch feature extraction module includes a time-frequency feature extraction module. Please refer to... Figure 4 The time-frequency feature extraction module includes a smooth pseudo-Wigner-Willi distribution transformation unit, a time-frequency map preprocessing unit, a head convolution unit, a residual feature extraction unit, and a global average pooling unit.

[0058] Specifically, S13 also includes the following steps S131 to S135: S131. Receive complex in-phase orthogonal sequences using a smooth pseudo-Wigner-Willi distribution transform unit and output a time-frequency distribution diagram.

[0059] Among them, the smooth pseudo-Wigner-Willi distribution transform unit is used to convert the input complex in-phase orthogonal sequence into a time-frequency energy distribution map, which captures the two-dimensional high-resolution time-frequency joint features of the signal while suppressing cross-term interference.

[0060] The time-frequency distribution diagram is calculated according to formula (5): (5); in, This is a time-frequency distribution diagram. For discrete time points, For discrete frequency points, Given a complex, in-phase, orthogonal sequence as input. The time offset variable for the frequency smoothing window. This is half the length of the frequency smoothing window. For complex sequences Perform conjugate complex number operations. For real symmetric frequency smoothing window function, The time offset variable for the time smoothing window. Half the length of the time smoothing window. It is a real symmetric time-smoothing window function. This represents the number of points in the Discrete Fourier Transform. S132. The time-frequency distribution map is received by the time-frequency map preprocessing unit, and logarithmic compression, upsampling and normalization are performed on the time-frequency distribution map in sequence to output the preprocessed time-frequency distribution map.

[0061] The time-frequency map preprocessing unit is used to perform amplitude normalization and size clipping and regularization on the generated time-frequency energy distribution map, eliminate the energy reference bias of the background environment, and adapt to the input dimension of the subsequent network.

[0062] Specifically, the logarithmic compression is calculated according to formula (6): (6); in, The time-frequency distribution matrix is ​​logarithmically compressed. A small constant to prevent logarithmic divergence.

[0063] S133. Receive the preprocessed time-frequency distribution map using the head convolution unit and output the preliminary time-frequency feature map.

[0064] The head convolutional unit is used to extract initial features from the preprocessed time-frequency map using shallow two-dimensional convolutional kernels, and initially captures the time-frequency joint texture and local edge assets in the time-frequency plane.

[0065] Specifically, the head convolutional unit comprises a two-dimensional convolutional layer, a batch normalization layer, a linear rectified activation layer, and a max pooling layer connected in sequence; wherein, the two-dimensional convolutional layer has 1 input channel, 64 output channels, and a kernel size of 7. 7, step size is 2, padding is 3.

[0066] S134. The residual feature extraction unit receives the preliminary time-frequency feature map, extracts higher-level two-dimensional texture features, frequency evolution trajectory features and overall energy distribution features from the preliminary time-frequency feature map layer by layer, and outputs a two-dimensional feature map.

[0067] Among them, the residual feature extraction unit is used to mine deep spatial features of shallow time-frequency textures by utilizing cascaded residual skip structures, while preserving low-order details and capturing high-order cross-domain intra-pulse modulation features.

[0068] Specifically, the residual feature extraction unit includes four residual layers, each of which includes two basic residual blocks; the number of output channels of each residual layer is 64, 128, 256 and 512 respectively, the step size of the first basic residual block of the second to fourth residual layers is 2, and the step size of the remaining basic residual blocks is 1; each basic residual block is equipped with a channel attention unit and a spatial attention unit.

[0069] The intermediate feature map is generated in the cascaded feature flow inside the basic residual block. Inside any of the basic residual blocks, the input data is extracted by the previous two-dimensional convolutional layer in the block and outputs a three-dimensional tensor containing local time-frequency texture features. This three-dimensional tensor is then used as the intermediate feature map and input into the channel attention unit set inside the block.

[0070] Let the intermediate feature map be The output of the channel attention unit is calculated according to formula (7): (7); in, Represents the channel attention weight vector. This represents a multilayer perceptron. This represents the Sigmoid activation function. For global average pooling operators, This is the global max pooling operator.

[0071] The output of the spatial attention unit is calculated according to formula (8): (8); in, This represents the new intermediate feature map after channel attention adjustment (i.e. ), This represents the concatenation operator for channel dimensions. This indicates that the kernel size is 7. 7. Two-dimensional convolution operation, This represents the spatial attention weight matrix.

[0072] S135. The two-dimensional feature map is received using a global average pooling unit, aggregated into a 512-dimensional time-frequency feature vector, and then output.

[0073] The global average pooling unit is used to perform spatial global average compression on each channel of the deep time-frequency feature map, which collapses and projects the high-dimensional spatial features to output the final time-frequency feature vector.

[0074] The time-frequency feature vector refers to the multidimensional feature sequence to be normalized, which is output by the time-frequency feature extraction module after performing joint time and frequency analysis, deep spatial texture mining and spatial dimension compression on the input complex in-phase orthogonal sequence. This sequence integrates two-dimensional time-frequency energy distribution features and has not yet undergone cross-domain scale and distribution alignment.

[0075] S14. Map the time-series feature vector and the time-frequency feature vector to a unified common feature space for scale and distribution alignment to obtain the time-series and time-frequency features of the radar signal.

[0076] Specifically, the timing features output from the timing branch are input into the timing alignment unit to obtain the processed timing features; the time-frequency features output from the time-frequency branch are input into the time-frequency alignment unit to obtain the processed time-frequency features. The processed timing features and the processed time-frequency features are then mapped to a unified common feature space.

[0077] The timing alignment unit and the time-frequency alignment unit both include a linear layer, a layer normalization layer and a linear rectification activation layer connected in sequence; the output dimension of the timing alignment unit is 128 and the output dimension of the time-frequency alignment unit is 128.

[0078] For example, the aligned time-series characteristics are calculated according to formula (9), and the aligned time-frequency characteristics are calculated according to formula (10).

[0079] (9); (10); in, As a time series feature, It has time-frequency characteristics. It is a linear rectified activation function. The original temporal feature vectors to be aligned. The original time-frequency feature vector to be aligned. This is a layer normalization layer.

[0080] It should be noted that S11 to S14 convert the original radar signal into a complex in-phase orthogonal sequence, providing complete original time-domain digital information for subsequent processing; then, the time-series and time-frequency feature extraction modules are used for parallel processing to extract the time-series feature vector and time-frequency feature vector within the signal pulse, respectively, realizing complementary acquisition of dual-domain information.

[0081] By mapping two heterogeneous vectors to a unified common feature space for scale and distribution alignment, the mismatch of feature dimensions (scale) and the bias of statistical intervals (distribution) caused by channel heterogeneity in time series and time frequency features are smoothed out. This blocks the traditional problem of heterogeneous feature space distribution mismatch at the source of features and provides a heterogeneous feature base with high comparability, balance and purity.

[0082] S2. Perform difference operations based on time-series features and time-frequency features to obtain difference features; perform dot product operations based on time-series features and time-frequency features to obtain related features; concatenate the time-series features, time-frequency features, difference features, and related features to obtain an interactively enhanced feature vector.

[0083] Among them, the difference feature is used to explicitly capture the complementary heterogeneous information of temporal features and time-frequency features at the same dimensional position.

[0084] Related features are used to explicitly extract the shared information of consistency between time-series features and time-frequency features in spatiotemporal resonance.

[0085] In one possible implementation, step S2 involves performing a difference operation based on time-series features and time-frequency features to obtain the difference features, including the following steps: S21. Calculate the element-level absolute difference between the time series features and the time-frequency features to obtain the absolute difference vector; use the absolute difference vector as the difference feature.

[0086] Specifically, the feature interaction enhancement module includes a difference feature construction unit, which inputs the aligned temporal features and time-frequency features into the difference feature construction unit to calculate the difference features of the dual-branch features in the common space.

[0087] The difference characteristics are calculated according to formula (11): (11); in, As a distinguishing feature, As a time series feature, It has time-frequency characteristics.

[0088] Calculating the element-level absolute difference between time-series features and time-frequency features can explicitly capture the microscopic inconsistencies between radar signals in the time-domain phase transition and time-frequency energy trajectory, preventing interference from heterogeneous data in traditional direct splicing.

[0089] Obtaining the absolute difference vector and extracting it as a difference feature allows this embodiment of the invention to no longer treat the spatial heterogeneity between the two domains as a "noise defect" hindering fusion, but rather actively transform it into a high-order complementary representation asset. This enriches the expressive completeness of multimodal features without introducing additional external noise.

[0090] In one possible implementation, step S2 involves performing a dot product operation based on time-series features and time-frequency features to obtain relevant features, including the following steps: S22. Perform element-wise multiplication of the time-series features and the time-frequency features to obtain an element-level dot product vector representing the consistency information between the time-series features and the time-frequency features; use the element-level dot product vector as the relevant feature.

[0091] The feature interaction enhancement module includes a related feature construction unit. The core function of the related feature construction unit is to provide a channel for extracting common features and filtering noise across multiple modalities. Specifically, during execution, the related feature construction unit directly receives the temporal and time-frequency features from the previous stage after scale and distribution alignment, and explicitly calculates and outputs the element-level dot product vector in the common space by executing the mathematical dot product operator of element-wise multiplication, thus solidifying it into related features.

[0092] Specifically, the temporal and time-frequency features are input into the relevant feature construction unit to calculate the relevant features of the dual-branch features in the common space.

[0093] The relevant features are calculated according to formula (12): (12); in, For relevant features, As a time series feature, It has time-frequency characteristics.

[0094] In step S22, since the preceding step S1 has already normalized the temporal and time-frequency features into a "unified common feature space," performing "element-by-element multiplication (Hadamard product)" is mathematically and physically justifiable. Essentially, it's a cross-domain joint attention mechanism. When two domains simultaneously possess high energy (i.e., strong consistency information) in a corresponding dimension, the product is exponentially amplified; conversely, when a domain experiences noise interference in that dimension due to spatial mismatch, multiplication effectively weakens it. This explicitly extracts the shared feature assets of the two domains at the underlying level, ensuring they are not overwhelmed by each other.

[0095] By using element-wise dot product vectors as relevant feature steps, the extracted cross-domain consistency information is solidified into independent "relevant features." This allows the invention to not only utilize single-domain specific information but also explicitly integrate the commonalities of multimodal intersections at the feature level. This overcomes the severe bias and mismatch in heterogeneous data spatial distribution caused by simple concatenation in traditional methods, fully extracting common representations for low-probability-of-interception radar signals and providing feature support for subsequent classification decisions.

[0096] In one possible implementation, step S2 involves concatenating time-series features, time-frequency features, difference features, and correlation features to obtain an interactive enhancement feature vector, including the following steps: S23. Concatenate and stitch together the time-series features, time-frequency features, difference features, and correlation features along the channel dimension to obtain the interactive enhancement feature vector.

[0097] Among them, the interactive enhancement feature vector refers to the multi-dimensional, three-dimensional, comprehensive feature representation vector generated by combining temporal features, time-frequency features, difference features, and related features through a channel dimension cascading splicing mechanism. This vector combines single-domain specificity with cross-domain explicit interactive assets.

[0098] Specifically, the interactive enhancement feature vector is calculated according to formula (13): (13); in, As a distinguishing feature, As a distinguishing feature, As a time series feature, It has time-frequency characteristics. This indicates a feature splicing operation.

[0099] In step S23, the problem of single feature representation dimension and mutual submersion of heterogeneous information caused by simply concatenating the original features in the time and frequency domains in radar signal recognition is alleviated. Based on the scale and distribution alignment already achieved in the previous stage of this invention, by performing element concatenation and splicing in the channel dimension, single-domain specific assets (time-series features, time-frequency features) are combined with cross-domain interactive assets (differential features reflecting micro-inconsistencies, and consistent correlation features reflecting cross-domain resonance), thus constructing a multi-dimensional three-dimensional feature space.

[0100] While preserving the original time-domain and frequency-domain micro-details, this invention explicitly introduces complementary and common features, enriching the representation assets of radar signal intra-pulse modulation. It also regularizes and alleviates the spatial distribution mismatch problem of heterogeneous data in the channel dimension, thus providing a discriminative feature basis for subsequent classification networks to make category decisions when facing radar signals with complex or specific intra-pulse modulation patterns.

[0101] S3. Input the interactive enhancement feature vector into the dynamic weight generation network to obtain the branch fusion weights corresponding to the interactive enhancement feature vector; use the branch fusion weights to perform weighted summation on the time-series features and time-frequency features respectively, and perform residual mapping on the weighted summation results to obtain the final fused features.

[0102] The dynamic weight generation network, based on the feature distribution characteristics of the current input sample, adaptively predicts and outputs the temporal branch weights and time-frequency branch weights to characterize the confidence of different feature channels through progressive dimensionality reduction, nonlinear excitation, and normalization.

[0103] Residual mapping refers to a feature refinement mechanism that uses a residual refinement network to perform deep nonlinear extraction on the weighted summation of the coarse fused feature stream, and introduces a cross-layer identity mapping path to perform element-level summation with the coarse fused feature stream, so as to suppress redundant background noise while preserving the original skeleton information.

[0104] Specifically, the dynamic weight generation network consists of a fully connected layer, a linear rectified activation layer, another fully connected layer, and a soft maximum normalization layer connected in sequence; the input dimension is 512 and the output dimension is 2.

[0105] In one possible implementation, step S3 involves inputting the interactive enhancement feature vector into a dynamic weight generation network to obtain the branch fusion weights corresponding to the interactive enhancement feature vector, including the following steps S31~S32: S31. Perform two-level nonlinear feature reduction and mapping on the interaction enhancement feature vector to obtain the branch score vector.

[0106] Among them, the branch score vector refers to the two-dimensional numerical representation vector output in the low-dimensional score space after the dynamic weight generation network performs two-level fully connected linear mapping and nonlinear activation on the input interactive enhancement feature vector. It is used to quantitatively indicate the credibility or quality weight of the current input sample in the time-series channel and the time-frequency channel.

[0107] In one possible implementation, step S31 involves performing two-level nonlinear feature reduction and mapping on the interactively enhanced feature vector to obtain the branch score vector, including the following steps S311~S313: S311. Using the first weight matrix and the first bias term, perform a first-level fully connected linear mapping on the interactive enhancement feature vector to obtain the first-level mapped feature vector.

[0108] The first weight matrix includes a matrix of trainable coefficient parameters configured in the first-level fully connected network layer to perform matrix multiplication with the input interactive enhancement feature vector, so as to perform preliminary compression and spatial basis reconstruction of multidimensional cascaded features in the spatial dimension.

[0109] The first bias term includes a training bias parameter vector configured in the first-level fully connected network layer and added element-wise to the intermediate result after matrix multiplication. This vector is used to perform a linear translation of the feature distribution in the vector space and enhance the robustness of the network fitting.

[0110] The first-level fully connected linear mapping refers to the mathematical linear transformation process that uses the first weight matrix and the first bias term to perform an affine transformation operation on the input interactive enhancement feature vector, decoupling the high-dimensional cross-domain spliced ​​features and projecting them onto a set low-dimensional transition feature space.

[0111] Specifically, let the interactive enhancement feature vector be... The first-level mapping feature vector is then calculated according to formula (14): (14); in, This is the first weight matrix. To enhance the interactive feature vector, For the first bias term, This is the first-level mapping feature vector.

[0112] S312. Use the linear rectified activation function to nonlinearly excite the first-level mapped feature vector and output the intermediate dimensionality reduction feature. The linear rectified activation function is used to introduce nonlinear excitation. Specifically, the linear rectified activation function includes the ReLU activation function.

[0113] S313. Using the second weight matrix and the second bias term, perform a second-level fully connected linear mapping on the intermediate dimensionality-reduced features to obtain the branch score vector.

[0114] The second weight matrix is ​​used to further compress the feature stream after nonlinear excitation and project it onto the training coefficient parameter matrix of the corresponding channel two-dimensional scoring space.

[0115] The second bias term includes a training bias parameter vector configured in the second-level fully connected network layer and element-wise superimposed onto the intermediate result after the second-level matrix multiplication operation. It is used to perform spatial linear translation on the scoring output and adjust the initial confidence benchmark of each channel.

[0116] The branch score vector is used to receive the low-dimensional feature stream and quantitatively characterize the real-time representation quality and feature confidence of the current input sample in the time-series feature channel and the time-frequency feature channel in the score space in the form of explicit numerical values.

[0117] Specifically, the branch score vector is calculated according to formula (15): (15); in, For branch score vectors, This is the second weight matrix. For the second bias term, This is the ReLU activation function. In steps S311 to S313, when performing the first-level mapping, by introducing the first weight matrix and the first bias term to perform a fully connected operation, the spatial dimension compression and feature recombination of the multi-dimensional interactive enhancement feature vector after the previous channel splicing can be performed to obtain the first-level mapping feature vector in the transition form.

[0118] Subsequently, a linear rectified activation function is used to nonlinearly excite the first-level mapped feature vector. Since this activation function has one-sided suppression properties, it can actively suppress negative redundant components less than zero in the mathematical space. While introducing nonlinear expressive power, it also achieves feature sparsity and initial noise filtering, producing intermediate dimensionality-reduced features with higher expressive purity.

[0119] Finally, the second-level fully connected mapping is performed on the intermediate dimensionality-reduced features using the second weight matrix and the second bias term, which regularizes the features and projects them onto the corresponding scoring space to obtain low-dimensional branch scoring vectors.

[0120] This specific multi-level composite mapping mechanism eliminates the ambiguity of the network structure at the implementation level. Through the synergistic effect of cascaded mapping and unilateral inhibition activation, the model can resolve the nonlinear mapping link between high-dimensional interactive features and the final branch score in a progressive dimensionality reduction manner. While controlling the number of network parameters and computational complexity, it enhances the model's sensitivity to capturing changes in the quality of input samples, thus laying the foundation for low-dimensional feature flow assets that conform to physical and mathematical logic for subsequent normalization processing and dual-branch adaptive weight generation.

[0121] S32. Normalize the branch scoring vector to generate time-series branch weights and time-frequency branch weights, and form the branch fusion weights by combining the time-series branch weights and time-frequency branch weights.

[0122] Among them, the temporal branch weights are used to dynamically adjust the feature expression intensity of the temporal channels as needed at the sample level, so as to adaptively retain or suppress the temporal phase evolution features of the current sample during the fusion stage.

[0123] The time-frequency branch weights are used to dynamically adjust the feature expression intensity of the time-frequency channels at the sample level as needed, thereby adaptively preserving or suppressing the two-dimensional time-frequency texture features of the current sample during the fusion stage.

[0124] Branch fusion weight refers to a two-dimensional coefficient vector composed of temporal branch weights and time-frequency branch weights, which is adaptively generated by a dynamic weight generation mechanism based on the real-time quality of the input samples. It is used to tune and quantify the fusion ratio of temporal and time-frequency channels at the sample level as needed.

[0125] Specifically, the branch fusion weight is calculated according to formula (16): (16); in, For time-series branch weights; The weights are for the time-frequency branches; and satisfy the following conditions: . For normalized exponential functions, This is the branch score vector.

[0126] In steps S31 to S32, two-level nonlinear feature reduction and mapping are performed on the cascaded interactive enhancement feature vectors of the previous stage. While realizing the dimensionality reduction representation of cross-domain joint assets, redundant components in the features are adaptively suppressed through nonlinear transformation, thereby obtaining a branch scoring vector with adaptive adjustment capability. Furthermore, by normalizing the branch scoring vector, temporal branch weights and time-frequency branch weights for the current input sample are generated, and then combined to form the branch fusion weights.

[0127] The aforementioned dynamic weight generation mechanism enables this invention to depart from the traditional static cascade mode, transforming the feature fusion process into a sample-level dynamic tuning mechanism. Without relying on manual prior setting of fusion ratios, it explicitly establishes a quantization link that dynamically adjusts the weight ratios of time-series and time-frequency channels based on the characteristics of the input samples. This alleviates local mismatches and random noise interference in heterogeneous feature spaces under complex environments, thus providing a weight foundation with sample-adaptive tuning capabilities for subsequent accurate weighted summation and residual refinement based on branch fusion weights.

[0128] In one possible implementation, step S3 uses branch fusion weights to perform weighted summation on the time-series features and time-frequency features respectively, and performs residual mapping on the weighted summation results to obtain the final fused features, including the following steps a1-a6: a1. Multiply the temporal branch weights and temporal features element-wise to obtain the first product result.

[0129] a2. Multiply the time-frequency branch weights and time-frequency features element-wise to obtain the second product result.

[0130] a3. Summing the first product result with the second product result yields a coarse fusion feature vector.

[0131] Among them, the coarse fusion feature vector is used to perform deep nonlinear extraction and noise purification in the subsequent input to the residual refining network, providing a multimodal mixed data stream with fully aligned spatial dimensions.

[0132] Specifically, the coarse fusion feature vector is calculated according to formula (17): (17); in, For time-series branch weights, For time-frequency branch weights, As a time series feature, It has time-frequency characteristics. This is a coarse-grained fusion feature vector.

[0133] a4. Input the coarse fusion feature vector into the residual refinement network and output the intermediate enhanced features.

[0134] The intermediate enhancement features are used to provide a refined feature stream that has been adaptively purified and freed from low signal-to-noise ratio redundant background noise, so as to facilitate subsequent linear mapping and construction of residual skip connections.

[0135] In one possible implementation, the residual refining network in step S36 includes: The network consists of a linear layer, a layer normalization layer, a linear rectified activation layer, and a random deactivation layer connected in sequence; the input and output dimensions of the residual refinement network are both 128.

[0136] The linear layer is used to perform matrix multiplication with the input coarse fused feature vector. While keeping the 128-dimensional feature scale unchanged, it performs cross-channel nonlinear spatial basis reconstruction and feature decoupling on the multi-channel mixed feature flow.

[0137] The layer normalization layer is used to perform statistical normalization and regularization of the mean and variance of the feature distribution output by the linear layer within the sample, adjusting the feature flow to a standard distribution range, thereby stabilizing the data flow inside the deep network and accelerating model convergence.

[0138] The linear rectified activation layer is used to perform a one-sided suppression nonlinear transformation on the normalized feature stream, adaptively filtering out non-positive redundant background noise with values ​​less than zero, thereby giving the residual refining network the ability to nonlinearly fit the intra-pulse modulation patterns of complex radar signals.

[0139] Random deactivation layers are used to randomly set the activation values ​​of some neurons to zero according to a set probability during the network training phase. This disrupts the co-adaptation between neurons, suppresses the model's over-reliance on specific noise features, and thus reduces the risk of overfitting during the refining process.

[0140] In the identity mapping pathway, the combination of linear layers and layer normalization layers enables nonlinear spatial mapping of the coarse fusion feature flow and performs normalization statistical regularization on the feature distribution within the layer. Subsequently, the one-sided suppression characteristic of the linear rectified activation layer (ReLU) is used to adaptively filter out negative redundant noise less than zero. Combined with the random node suppression of the random deactivation layer (Dropout) during the training phase, the overfitting phenomenon of the feature flow during the refinement process is alleviated.

[0141] By uniformly limiting the input and output dimensions to 128, the technical foundation for matching is laid for subsequent steps, ensuring lossless alignment of feature information during transmission.

[0142] a5. Perform linear mapping on the intermediate enhanced features to obtain deep refined features.

[0143] The deep refined features include high-order feature vectors that are perfectly aligned with the original coarse fused feature vectors in both channel dimension and numerical scale, output by the linear mapping layer after performing spatial topology adjustments on the intermediate enhanced features. These vectors are primarily used to provide high-fidelity intra-pulse modulation fine-grained information purified and normalized by multiple deep networks at the feature level, enabling cross-layer complementary advantages with the shallow skeleton features.

[0144] a6. Perform element-wise summation on the coarse fused feature vector and the deep refined feature vector to obtain the final fused feature.

[0145] The final fusion feature includes a three-dimensional comprehensive representation vector that integrates multimodal dynamic channel weights, output by element-wise summation of the original coarse fusion feature vector and the deeply refined deep features after deep purification by the residual jump connection pathway.

[0146] Specifically, the final fusion feature is calculated according to formula (18): (18); in, For the final fusion feature, Represents the residual refining mapping function; This is the learnable scaling factor.

[0147] In steps a1-a6, during channel-level dynamic adjustment, the weights of the temporal branch and the weights of the time-frequency branch are element-wise multiplied with their corresponding temporal and time-frequency features and then summed. This method mathematically implements the sample-level allocation ratio generated in the previous stage, realizing the physical intention of dynamically adjusting the proportion of temporal phase and time-frequency texture channels according to the quality of input samples, and initially producing a coarse fusion feature vector that aggregates cross-domain exchange assets. Subsequently, the coarse fused feature vector is input into the residual refinement network, and intermediate enhanced features are output after nonlinear transformation. These enhanced features are then obtained through linear mapping to obtain deep refined features. Utilizing the feature extraction capabilities of the network layers, deep nonlinear extraction and regularization are performed on the coarse feature stream after multi-channel mixing. This adaptively suppresses redundant background noise left over from low signal-to-noise ratio environments or mismatches in the previous feature space distribution, achieving feature purification. Finally, a residual skip connection path is constructed by performing element-wise summation of the initial coarse fusion feature vector and the refined deep refinement feature vector, enabling the initial coarse fusion feature information to bypass the deep refinement network and be directly transmitted to the output.

[0148] Through steps a1 to a6, the ambiguity of the fusion algorithm is eliminated at the implementation level. By combining dynamic channel-level tuning with a cross-layer residual identity mapping mechanism, on the one hand, the problem of feature gradient vanishing or loss of key intra-pulse modulation details that is easily caused by increasing the number of layers in traditional deep neural networks is alleviated during deep refinement, while preserving the basic skeleton information of the original time-frequency and temporal dimensions; on the other hand, the distribution bias and spatial mismatch bias of heterogeneous feature streams are eliminated, thereby providing robust final fusion features for the stable category decision output of the terminal classification network.

[0149] S4. Input the final fused features into the pre-trained classification network and output the intra-pulse modulation category corresponding to the radar signal.

[0150] The classification network receives high-order final fusion features and internally reorganizes the abstract feature space and projects it onto the category decision scoring space through multi-layer linear mapping, distribution normalization, nonlinear activation and overfitting suppression. Finally, it adaptively outputs the intra-pulse modulation category corresponding to the radar signal.

[0151] In one possible implementation, the classification network includes a fully connected layer, a batch normalization layer, a linear rectified activation layer, a random deactivation layer, and a fully connected output layer connected in sequence; wherein, the output dimension of the fully connected layer is 256, the output dimension of the fully connected output layer is 15, and the output dimension of the fully connected output layer corresponds to 15 categories of radar signal intra-pulse modulation.

[0152] The fully connected layer is used to perform matrix multiplication with the final fused features of the input, projecting and decoupling the 128-dimensional feature vector to a 256-dimensional high-dimensional hidden space to achieve cross-channel nonlinear fusion of low-order mixed features and feature space reorganization.

[0153] The batch normalization layer performs statistical normalization correction on the mean and variance of the feature distribution output by the fully connected layer within the training batch, forcing the data flow to be constrained to a relatively stable central distribution interval, thereby alleviating the internal covariate shift of the input data in deep transmission and accelerating network convergence.

[0154] The linear rectified activation layer is used to perform a one-sided suppression nonlinear transformation on the normalized hidden layer features, forcibly clearing the negative activation values ​​to filter out the remaining non-positive redundant noise, thereby giving the classification terminal the ability to make nonlinear decisions and boundary divisions for 15 complex modulation patterns.

[0155] The random deactivation layer is used to randomly shut down the feature transmission of some neuron nodes according to a set probability during the training phase of the classification terminal, thereby destroying the co-adaptive dependency between feature channels and improving the generalization performance and noise robustness of the classifier when facing low signal-to-noise ratio or unknown noise interference.

[0156] The fully connected output layer is used to linearly map the 256-dimensional refined hidden features to a 15-dimensional category scoring space using the output weight matrix and bias terms. It directly quantifies the prediction confidence of the current input sample in 15 specific radar signal pulse modulation categories with explicit numerical scores, thus completing the entire closed loop of the method from "high-order features" to "accurate 15-class output at the terminal" at the implementation level. Specifically, first, the classification network receives the 128-dimensional final fused features from the previous stage output. Inside the network terminal, the final fused features are first input to a fully connected layer for matrix multiplication, and the feature vectors are projected and decoupled to a 256-dimensional high-dimensional hidden space, completing the nonlinear recombination of the feature space across channels.

[0157] Subsequently, the recombined 256-dimensional data stream enters the batch normalization layer, where it undergoes statistical normalization correction of the mean and variance within the training batch. The feature distribution is forced to a stable central distribution interval to mitigate internal covariate bias.

[0158] Then, the normalized hidden layer features are subjected to a nonlinear transformation of one-sided inhibition through a linear rectified activation layer to clear the negative activation values ​​to filter out the remaining non-positive redundant noise; and at the same time, the feature transmission of some neuron nodes is randomly shut down according to a set probability using a random deactivation layer to destroy the co-adaptation dependency between channels and suppress the risk of overfitting.

[0159] Finally, the purified and overfit-suppressed 256-dimensional hidden features are input to the fully connected output layer. A linear mapping is performed using the output weight matrix and bias term to reduce the dimensionality of the features and project them into a 15-dimensional class scoring space. The 15 explicit numerical scores in this space directly quantify the prediction confidence of the current input sample in 15 specific radar signal intra-pulse modulation categories. Based on the index label corresponding to the highest score, the network adaptively outputs the final intra-pulse modulation category of the radar signal, completing the entire classification process.

[0160] In this embodiment of the invention, a fully connected layer is used to project the final fused features from the previous stage into a 256-dimensional hidden layer space, achieving nonlinear recombination of the feature space. Combined with a batch normalization layer, the statistical distribution within the channel is corrected, mitigating the distribution bias of the input data during deep propagation. Subsequently, through the synergistic effect of a linear rectified activation layer and a random deactivation layer, negative redundant noise is filtered out while reducing the risk of network overfitting.

[0161] Finally, by using a fully connected output layer to linearly map the hidden layer features to a 15-dimensional category scoring space, the precise correlation between the feature space and 15 specific radar signal intra-pulse modulation categories was achieved. At the implementation level, the entire process loop from "feature extraction and regularization" to "precise output of 15 categories at the terminal" was completed.

[0162] It should be noted that in steps S1 to S4, the traditional radar signal intra-pulse modulation identification suffers from a feature spatial distribution mismatch problem because the time-series features of the one-dimensional network and the time-frequency features of the two-dimensional network have different numerical ranges and mean-variance mismatch.

[0163] Step S1, as the first step of this invention, introduces a "scale and distribution alignment mechanism" at the source of feature extraction. This mechanism projects the temporal features of one-dimensional phase evolution information and the time-frequency features of two-dimensional energy texture information onto the same unified common feature space. This avoids discrepancies in the numerical ranges and mean-variance mismatch between the temporal features of the one-dimensional network and the time-frequency features of the two-dimensional network. Consequently, it reduces the feature space distribution mismatch problem that occurs in traditional radar signal intra-pulse modulation recognition and improves the accuracy of radar signal intra-pulse modulation recognition.

[0164] Step S2 constructs a multi-dimensional, three-dimensional interactive enhanced feature vector by concatenating single-domain specific features (time-series and time-frequency features) with cross-domain interactive features (difference and correlation features). Since the previous stage has already achieved scale and distribution alignment, this step can accurately realize deep explicit interaction of multimodal features, maximizing the use and preserving the complete representation of radar signal intra-pulse modulation. This avoids the forced submersion and redundant interference of heterogeneous multi-source information during fusion, further reducing the degree of feature space distribution mismatch, thereby improving the accuracy of radar signal intra-pulse modulation recognition.

[0165] Step S3 achieves secondary elastic regularization and noise suppression for the feature space distribution mismatch problem. The model no longer mechanically and rigidly allocates the fusion ratio of each representation domain, but dynamically adjusts the weight ratio of timing and time-frequency channels based on the real-time quality of different radar samples, achieving sample-level on-demand fusion. Simultaneously, through the residual mapping path, redundant noise remaining due to signal-to-noise ratio degradation or feature space mismatch is filtered out, extracting pure feature representations that aggregate cross-domain deep exchange assets. This eliminates the spatial distribution mismatch phenomenon of heterogeneous multimodal data and enhances the overall noise robustness of the solution.

[0166] Step S4 inputs the final fused features, refined through spatial adaptive dynamic fusion and residual mapping, into the pre-trained classification network. The classification network directly maps and determines the intra-pulse modulation category corresponding to the original radar signal and outputs the classification network. The classification network can make extremely sensitive and high-confidence decisions for various complex modern radar patterns with low intercept probability, solving the technical bottleneck of low radar signal recognition accuracy caused by feature space distribution mismatch in traditional methods based on a single representation domain or simple stitching.

[0167] This invention extracts the temporal and time-frequency features of radar signals, and then performs difference and dot product operations on these features to obtain difference and correlation features. These features are then concatenated to obtain an interactively enhanced feature vector. This avoids the problem of existing methods focusing on a single representation domain, thus improving feature utilization, reducing the impact of feature space distribution mismatch, and ultimately enhancing the accuracy of radar signal recognition.

[0168] Simultaneously, this invention inputs the interactive enhancement feature vector into a dynamic weight generation network to obtain the branch fusion weights of the interactive enhancement feature vector. The branch fusion weights are then used to perform weighted summation on the temporal features and the time-frequency features respectively, and residual mapping is applied to the weighted summation results to obtain the final fused features. This avoids the problem of existing methods simply concatenating temporal and time-frequency features, thereby solving the feature space distribution mismatch problem and improving the recognition accuracy of radar signals.

[0169] Secondly, to verify the effectiveness of the radar signal intra-pulse modulation recognition method based on dual-branch feature extraction and adaptive interactive fusion proposed in this invention, experiments were conducted using a simulation dataset containing fifteen typical radar intra-pulse modulation signals. The simulation experiments for radar signal intra-pulse modulation recognition are described below. 1. Simulation conditions The dataset used in this study is constructed based on a mathematical model of typical radar intra-pulse modulation signals, covering a total of 15 types of intra-pulse modulation signals, specifically including continuous wave, linear frequency modulation, sinusoidal frequency modulation, second-order frequency modulation, phase shift keying, frequency shift keying, Frank code, P1 code, P2 code, P3 code, P4 code, T1(n) code, T2(n) code, T3(n) code, and T4(n) code.

[0170] To ensure the samples cover typical feature variations under different modulation parameters, parameters such as carrier frequency, bandwidth, code length, phase basis, symbol length, frequency interval, and number of segments for various signals were randomly set during dataset construction. Specific simulation parameter settings are shown in Table 1. The channel environment used was additive white Gaussian noise, with a signal-to-noise ratio (SNR) range of -14dB to 10dB. The training set was constructed at 2dB intervals to control the training scale; the validation and test sets were constructed at 1dB intervals to verify the method's recognition performance under different SNR conditions. For each modulation type, 500 training samples were generated at each training SNR, and 100 validation samples and 100 test samples were generated at each validation and test SNR.

[0171] Table 1

[0172] 2. Experimental Setup The experimental environment configuration and training hyperparameter settings used in this invention are shown in Table 2.

[0173] Table 2

[0174] To verify the role of each module in the proposed dual-branch feature extraction and adaptive interactive fusion method, an ablation experiment was conducted. Comparative models were constructed, including only the temporal feature extraction branch, only the time-frequency feature extraction branch, and models with the adaptive fusion module and feature interaction enhancement module removed, respectively. These models were then compared with the complete model. The ablation experiment settings of this embodiment are shown in Table 3.

[0175] Table 3

[0176] 3. Results Analysis Depend on Figure 5 Analysis shows that the proposed method exhibits good recognition performance under various test conditions, indicating that it can stably complete the intra-pulse modulation category determination of radar signals. This result demonstrates that parallel feature extraction via temporal and time-frequency branches can compensate for the shortcomings of a single representation method in terms of information coverage, thereby obtaining a more complete joint representation.

[0177] Depend on Figure 6 Analysis shows that the present invention has good classification performance in most modulation categories. Although some structurally similar modulation types are still somewhat confused, the overall proportion of the main diagonal lines in each category is relatively high, indicating that the present invention can effectively extract key differences between different modulation methods and achieve relatively stable category differentiation.

[0178] Depend on Figure 7 Analysis shows that the proposed method outperforms the comparison models in recognition accuracy under different signal-to-noise ratio conditions. The classification accuracy is close to 100% above 0dB; below -6dB, only the accuracy of the time-series branch model 1 drops below 60%, while the single-time-frequency branch model 2 is slightly better but still inferior to the complete method of this invention; models 3 and 4, with the adaptive fusion or feature interaction modules removed, perform close to the single-time-frequency branch, indicating that simple combinations cannot fully utilize the complementary features of the two branches; the complete method exhibits high accuracy across the entire signal-to-noise ratio range, achieving effective fusion and joint discrimination of time-series and time-frequency features.

[0179] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0180] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A radar signal intra-pulse modulation identification method based on branch feature extraction and fusion, characterized in that, include: Acquire radar signals and extract time-series and time-frequency features aligned with the scale and distribution of the radar signals; Difference operations are performed on the time-series features and the time-frequency features to obtain difference features; dot product operations are performed on the time-series features and the time-frequency features to obtain correlation features; the time-series features, the time-frequency features, the difference features, and the correlation features are concatenated to obtain an interaction-enhanced feature vector; The interaction enhancement feature vector is input into the dynamic weight generation network to obtain the branch fusion weights corresponding to the interaction enhancement feature vector. The time-series features and the time-frequency features are weighted and summed using the branch fusion weights, and the weighted summation results are residual mapped to obtain the final fused features. The final fused features are input into a pre-trained classification network, which outputs the intra-pulse modulation category corresponding to the radar signal.

2. The radar signal intra-pulse modulation identification method based on branch feature extraction and fusion according to claim 1, characterized in that, Acquire radar signals, and extract the scale- and distribution-aligned temporal and time-frequency features of the radar signals, including: Acquire radar signals and convert them into complex in-phase orthogonal sequences; The complex in-phase orthogonal sequence is input into the time series feature extraction module to obtain the time series feature vector; The complex in-phase orthogonal sequence is input into the time-frequency feature extraction module to obtain the time-frequency feature vector; The temporal feature vector and the time-frequency feature vector are mapped to a unified common feature space for scale and distribution alignment to obtain the temporal and time-frequency features of the radar signal.

3. The radar signal intra-pulse modulation identification method based on branch feature extraction and fusion according to claim 1, characterized in that, Based on the time-series features and the time-frequency features, difference operations are performed to obtain difference features, including: Calculate the element-level absolute difference between the time-series feature and the time-frequency feature to obtain an absolute difference vector; use the absolute difference vector as the difference feature.

4. A radar signal intra-pulse modulation identification method based on branch feature extraction and fusion according to claim 1, characterized in that, Based on the time-series features and the time-frequency features, a dot product operation is performed to obtain relevant features, including: The time-series feature and the time-frequency feature are multiplied element-wise to obtain an element-level dot product vector representing the consistency information between the time-series feature and the time-frequency feature; the element-level dot product vector is used as the relevant feature.

5. A radar signal intra-pulse modulation identification method based on branch feature extraction and fusion according to claim 1, characterized in that, The time-series features, time-frequency features, difference features, and correlation features are concatenated to obtain an interactive enhancement feature vector, including: The time-series features, time-frequency features, difference features, and correlation features are concatenated and strung together element by element along the channel dimension to obtain the interaction-enhanced feature vector.

6. A radar signal intra-pulse modulation identification method based on branch feature extraction and fusion according to claim 1, characterized in that, The interaction enhancement feature vector is input into a dynamic weight generation network to obtain the branch fusion weights corresponding to the interaction enhancement feature vector, including: The interaction enhancement feature vector is subjected to two-level nonlinear feature dimensionality reduction and mapping to obtain the branch score vector; The branch scoring vector is normalized to generate time-series branch weights and time-frequency branch weights, and the branch fusion weights are composed of the time-series branch weights and the time-frequency branch weights.

7. The method according to claim 6, characterized in that, The interaction enhancement feature vector is subjected to two-level nonlinear feature dimensionality reduction and mapping to obtain the branch scoring vector, including: The interaction enhancement feature vector is subjected to a first-level fully connected linear mapping using the first weight matrix and the first bias term to obtain the first-level mapped feature vector; The first-level mapped feature vector is nonlinearly activated using a linear rectified activation function to output intermediate dimensionality-reduced features. The branch score vector is obtained by performing a second-level fully connected linear mapping on the intermediate dimensionality-reduced features using the second weight matrix and the second bias term.

8. A radar signal intra-pulse modulation identification method based on branch feature extraction and fusion according to claim 6, characterized in that, The time-series features and the time-frequency features are weighted and summed using the branch fusion weights, and the weighted summation results are residual mapped to obtain the final fused features, including: The element-wise product of the temporal branch weights and the temporal features is performed to obtain the first product result. The time-frequency branch weights are element-wise multiplied with the time-frequency features to obtain a second product result. The first product result and the second product result are summed to obtain the coarse fusion feature vector; The coarse fused feature vector is input into the residual refinement network, which outputs intermediate enhanced features. Linear mapping is performed on the intermediate enhancement features to obtain deep refined features; The coarse fused feature vector and the deep refined feature vector are summed element-wise to obtain the final fused feature.

9. A radar signal intra-pulse modulation identification method based on branch feature extraction and fusion according to claim 8, characterized in that, The residual refining network includes: The residual refinement network consists of a linear layer, a layer normalization layer, a linear rectified activation layer, and a random deactivation layer connected in sequence; wherein the input dimension and output dimension of the residual refinement network are both 128.

10. A radar signal intra-pulse modulation identification method based on branch feature extraction and fusion according to claim 1, characterized in that, The classification network includes: The system consists of a fully connected layer, a batch normalization layer, a linear rectification activation layer, a random deactivation layer, and a fully connected output layer, connected sequentially. The fully connected layer has an output dimension of 256, and the fully connected output layer has an output dimension of 15, which corresponds to 15 types of radar signal intra-pulse modulation categories.

Images