Pyramid multi-scale time sequence convolution network and corner reflector array identification method based on the network

Abstract

The present application belongs to the technical field of radar target recognition, and relates to a pyramid multi-scale time series convolution network and an array of corner reflectors recognition method based on the network. The pyramid multi-scale time series convolution network comprises a preprocessing unit, a feature extraction and fusion unit and a classification decision unit connected in sequence, the preprocessing unit is used for amplitude normalization and initial feature extraction on the HRRP sequence; the feature extraction and fusion unit comprises N cascaded pyramid blocks and a multi-scale pooling fusion subunit, each pyramid block uses a gradually expanding dilation coefficient to extract time series features of different scales, and the multi-scale pooling fusion subunit realizes multi-scale feature fusion through global average pooling and channel splicing; the classification decision unit outputs a target category. The present application captures micro, meso and macro time series features of the HRRP sequence through the pyramid multi-scale structure, and significantly improves the recognition accuracy and robustness of the ship target and the array of corner reflectors under low signal-to-clutter ratio.

Description

Technical Field

[0001] This invention relates to the field of radar target recognition technology, and in particular to a pyramid multi-scale temporal convolutional network and a method for recognizing corner reflector arrays based on this network. Background Technology

[0002] With the development of modern radar technology, high-resolution range profiles (HRRPs) can accurately reflect the fine geometric features of targets and have become a key basis for radar automatic target recognition (RATR). In maritime combat scenarios, corner reflector arrays, as an efficient passive jamming technology, can use multiple metallic reflective surfaces to reflect incident electromagnetic waves back along their original path, generating strong echoes at the radar receiver. Even corner reflectors with small geometric dimensions can exhibit a large radar cross section (RCS), and when multiple corner reflectors are arranged in an array, they can realistically simulate the HRRP of surface ships, severely interfering with the normal operation of radar systems.

[0003] Early HRRP time series feature extraction mainly relied on signal processing and statistical modeling methods. Signal processing methods used techniques such as Fourier transform and wavelet transform to map the one-dimensional HRRP to different physical domains to extract target features; statistical modeling methods parameterized the dynamic generation process of the HRRP sequence through statistical models, transforming feature extraction into model parameter estimation. Although these methods have clear physical meaning and strong interpretability, their feature representation ability and discrimination are limited, and they are poorly adaptable to complex interference and dynamic environments.

[0004] In recent years, Temporal Convolutional Networks (TCNs) have emerged as a novel sequence modeling architecture, offering new insights for radar temporal feature extraction. TCNs can efficiently capture long-range dependencies, support parallel computing, and exhibit efficient and stable training processes. However, standard TCNs employ a single-scale receptive field, failing to simultaneously capture the fine local scattering structure and global distribution patterns of HRRP under corner reflector array interference. Their multi-scale feature capture capabilities are insufficient, making it difficult to meet the accurate recognition requirements in complex scenarios. Especially in low signal-to-clutter ratio (SCR) scenarios, model recognition accuracy drops significantly, robustness is severely lacking, and it cannot meet the practical engineering requirements for anti-jamming of maritime radar.

[0005] In summary, existing temporal convolutional networks have shortcomings such as being unable to simultaneously take into account the local fine structure and global distribution pattern of HRRP, poor recognition performance at low signal-to-noise ratios, and low distinguishability between ship targets and corner reflector arrays. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a pyramid multi-scale temporal convolutional network and a corner reflector array recognition method based on this network. By configuring progressive expansion factors through three cascaded pyramid blocks, the network captures temporal features at three different scales in the HRRP sequence: microscopic instantaneous fluctuations of the scattering center, mesoscopic structural evolution, and macroscopic attitude changes. This significantly improves the ability to distinguish between ship targets and corner reflector arrays.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a pyramid multi-scale temporal convolutional network, comprising a preprocessing unit, a feature extraction and fusion unit, and a classification decision unit connected sequentially. The preprocessing unit is used to perform amplitude normalization and initial feature extraction on the input HRRP sequence raw data, and output the initial feature map; The feature extraction and fusion unit comprises N cascaded pyramid blocks and a multi-scale pooling fusion subunit, where N ≥ 2. The N pyramid blocks employ progressively increasing expansion coefficients. Each pyramid block includes a parallel main path and secondary paths. The main path is used to extract temporal features at the corresponding scale of the pyramid block, while the secondary paths are used to perform dimensional mapping on the input features to make them consistent with the output dimension of the main path. The multi-scale pooling fusion subunit is used to perform multi-scale fusion on the output feature maps of each pyramid block to obtain a multi-scale fused feature vector. The classification decision unit is used to process the multi-scale fused feature vector and output the target category recognition result.

[0008] As one possible implementation, N = 3, and the expansion coefficients of the three pyramid blocks are 1, 2 and 4, respectively, which are used to capture three types of temporal features of the scattering center in the HRRP sequence: microscopic instantaneous fluctuations, mesoscopic structural evolution and macroscopic attitude changes. The number of output channels of the three pyramid blocks are 64, 64 and 96, respectively.

[0009] As one possible implementation, in each pyramid block, the main path is used to sequentially perform two identical one-dimensional dilated convolutions, layer normalization, ReLU activation, and Dropout processing on the input features of the current pyramid block, with the Dropout rate ranging from 0.2 to 0.3, outputting the main path features; the secondary path is used to project the input features of the pyramid block to the same channel dimension as the output of the main path through a 1×1 convolutional layer, outputting the skip connection features; each pyramid block adds the main path features and skip connection features element-wise, and then performs ReLU activation and Dropout to obtain the output feature map of the current pyramid block.

[0010] As one possible implementation, the multi-scale pooling fusion subunit performs the following operations: Global average pooling is performed on the output feature map of each pyramid block along the temporal dimension to obtain N pooling vectors; By concatenating N pooling vectors along the channel dimension, a multi-scale fused feature vector is obtained.

[0011] As one possible implementation, the preprocessing unit includes a one-dimensional convolutional layer, a layer normalization layer, and a ReLU activation layer. The one-dimensional convolutional layer has 32 output channels, and the convolutional kernel uses same padding to keep the sequence length unchanged. The normalization process linearly maps the amplitude of each sample in the HRRP sequence to the interval [-1, 1].

[0012] As one possible implementation, the classification decision unit includes a feature refinement network and a classifier. The feature refinement network consists of two fully connected layers that map the multi-scale fused feature vectors into refined feature vectors. The classifier maps the refined feature vectors to class scores and outputs the probability distribution of each class through a Softmax function. The dropout rate after the fully connected layer of the feature refinement network is 0.3, and the dropout rate after the fully connected layer of the classifier is 0.2.

[0013] As one possible implementation method, the target recognition results include three categories: ship targets, corner reflector arrays, and single corner reflectors.

[0014] Secondly, the present invention provides a method for identifying corner reflector arrays based on a pyramid multi-scale temporal convolutional network as described in the first aspect, comprising: S1. Obtain the raw data of the HRRP sequence to be identified; S2. Input the raw data of the HRRP sequence to be identified into the trained pyramid multi-scale temporal convolutional network; S3. The preprocessing unit performs amplitude normalization and initial feature extraction on the input HRRP sequence raw data, and outputs the initial feature map; S4. The feature extraction and fusion unit performs multi-scale feature extraction and pooling fusion on the initial feature map, and outputs a multi-scale fused feature vector; S5. The classification decision unit processes the multi-scale fused feature vector and outputs the target category recognition result, which includes three categories: ship target, corner reflector array, and single corner reflector.

[0015] As one possible implementation, before step S1, the pyramid multi-scale temporal convolutional network is trained. The training steps include: obtaining a training sample set, wherein the corner reflector array samples in the training sample set cover a variety of deployment scenarios, including but not limited to single array, multi-group collaboration, and random scattering, and the target pose angle covers 0°~180°; and training the pyramid multi-scale temporal convolutional network using a joint loss function that is a weighted combination of cross-entropy loss and center loss.

[0016] As one possible implementation, the joint loss function is as follows: in, This represents the equilibrium hyperparameter, with a value range of 0.01 to 0.2; The cross-entropy loss is calculated using the following method: In the formula, Indicates batch size, Indicates the total number of categories. Indicates the first The true label of each sample Indicates the first The sample belongs to the first The predicted probability of a class; The central loss is represented and calculated using the following method: In the formula, Indicates the first The refined feature vector of each sample, Indicates the first Each sample corresponds to a feature center of its class.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention employs a pyramidal multi-scale temporal convolutional network architecture. Through three cascaded pyramidal blocks, it captures temporal features at three different scales in the HRRP sequence: microscopic instantaneous fluctuations of the scattering center, mesoscopic structural evolution, and macroscopic attitude changes. This overcomes the shortcomings of traditional single-scale TCNs, which have a limited receptive field and cannot simultaneously capture both local fine structure and global distribution patterns, significantly improving the ability to distinguish between ship targets and corner reflector arrays. Traditional single-scale feature-based identification methods (such as Fourier transform, wavelet transform, and statistical modeling) have limited feature representation capabilities. While standard TCNs can capture temporal dependencies, their single-scale receptive field cannot simultaneously capture microscopic scattering details and macroscopic attitude changes. This invention, through a progressive expansion factor configuration (expansion coefficients of 1, 2, and 4), achieves a gradual expansion of the receptive field from local to global, comprehensively characterizing the essential differences between ship targets and corner reflector arrays in scattering distribution, local structure, and global evolution patterns.

[0018] 2. This invention employs a multi-scale feature fusion strategy combining global average pooling and channel concatenation. It integrates features from three pyramid blocks at different scales along the channel dimension, achieving complementary enhancement of multi-granularity information. Simultaneously, the pyramid blocks utilize a dual-path residual structure, effectively addressing the gradient vanishing problem in deep network training and ensuring training stability and feature reuse efficiency. The residual structure, through the parallel design of the main and secondary paths, allows gradients to propagate directly via skip connections, preventing deep network degradation. Global average pooling compresses features along the temporal dimension, preserving key information at each scale. Channel concatenation then organically fuses multi-scale features, enabling the network to simultaneously utilize both microscopic details and macroscopic semantic information for comprehensive discrimination.

[0019] 3. This invention employs a joint loss function combining cross-entropy loss and center loss. While optimizing classification accuracy, the center loss constrains samples of the same class to cluster towards the class center in the feature space, while samples of different classes move away from each other. This effectively improves the intra-class aggregation and inter-class separability of the feature space, enhancing the model's feature discrimination ability. Cross-entropy loss measures the difference between the predicted probability distribution and the true label, optimizing the classification boundary; center loss maintains the feature center of each class, constraining sample features to move closer to their corresponding class center during training. The joint optimization of these two loss functions enables the model to obtain more discriminative feature representations while improving classification accuracy, further reducing the confusion probability between ship targets and corner reflector arrays.

[0020] 4. This invention exhibits excellent recognition robustness in low signal-to-clutter (SCR) scenarios, effectively resisting complex sea surface clutter interference. Its recognition accuracy across the entire SCR range surpasses existing comparative algorithms. Through comprehensive acquisition of multi-scale temporal features and discriminative optimization of the joint loss function, experiments show that under low SCR conditions of 0dB~4dB, the overall recognition accuracy is improved by at least 10 percentage points compared to comparative algorithms such as LSTM, SSM, and ResNet18. At a SCR of 20dB, the recognition accuracy for ships, corner reflector arrays, and individual corner reflectors reaches 96.7%, 97.5%, and 98.7%, respectively, meeting the practical engineering requirements for radar anti-jamming in complex naval warfare environments. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the pyramid multi-scale temporal convolutional network structure in an embodiment of the present invention; Figure 2 This is a schematic diagram of the internal structure of the pyramid block in an embodiment of the present invention; Figure 3 This is an example diagram of a simulation dataset sample in an embodiment of the present invention; Figure 4 These are schematic diagrams illustrating three typical deployment scenarios of corner reflector arrays in embodiments of the present invention; Figure 5 This is a comparison chart of the recognition performance of different algorithms under different signal-to-noise ratios in the embodiments of the present invention; Figure 6 This is a confusion matrix diagram of the recognition results of the pyramid multi-scale temporal convolutional network in an embodiment of the present invention at a signal-to-noise ratio of 20dB.

[0022] Figure Labels 1-Preprocessing unit, 10-One-dimensional convolutional layer, 11-Normalization layer, 12-ReLU activation layer; 2-Feature extraction and fusion unit, 20-Multi-scale pooling fusion sub-unit, 21-First pyramid block, 22-Second pyramid block, 23-Third pyramid block; 3-Classification decision unit, 30-Feature refinement network, 31-Classifier. Detailed Implementation

[0023] To facilitate a clear description of the technical solutions in the embodiments of the present invention, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. For example, the first threshold and the second threshold are merely used to distinguish different thresholds and do not limit their order. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" are not necessarily different.

[0024] It should be noted that in this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0025] In this invention, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one" or similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, "at least one of a, b, or c" can represent: a, b, c, a combination of a and b, a combination of a and c, a combination of b and c, or a, b, and c, where a, b, and c can be single or multiple.

[0026] The present invention aims to provide a pyramid multi-scale temporal convolutional network and a corner reflector array recognition method based on the network. By configuring progressive expansion factors through three cascaded pyramid blocks, the network captures three different types of temporal features at different scales in the HRRP sequence: microscopic instantaneous fluctuations of the scattering center, mesoscopic structural evolution, and macroscopic attitude changes. This overcomes the shortcomings of traditional single-scale TCNs, which have a single receptive field and cannot simultaneously take into account local fine structure and global distribution patterns.

[0027] In a first aspect, embodiments of the present invention provide a pyramid-shaped multi-scale temporal convolutional network, see [link to previous document]. Figure 1 It includes a preprocessing unit 1, a feature extraction and fusion unit 2, and a classification decision unit 3 connected in sequence. The preprocessing unit is used to perform amplitude normalization and initial feature extraction on the input HRRP sequence raw data, and output the initial feature map; See Figure 1As one possible implementation, the preprocessing unit 1 includes a one-dimensional convolutional layer 10, a layer normalization layer 11, and a ReLU activation layer 12. The one-dimensional convolutional layer 10 has 32 output channels, and the convolutional kernel uses same padding to keep the sequence length unchanged. The normalization process linearly maps the amplitude of each sample in the HRRP sequence to the interval [-1, 1].

[0028] The feature extraction and fusion unit 2 includes N cascaded pyramid blocks and a multi-scale pooling fusion subunit 20, where N ≥ 2. The N pyramid blocks use progressively increasing expansion coefficients. Each pyramid block includes a parallel main path and secondary paths. The main path is used to extract the temporal features of the corresponding scale of the pyramid block, and the secondary paths are used to perform dimensional mapping on the input features to make them consistent with the output dimension of the main path. The multi-scale pooling fusion subunit is used to perform multi-scale fusion on the output feature maps of each pyramid block to obtain a multi-scale fused feature vector. See Figure 1 As one possible implementation, N = 3, with 3 pyramid blocks: the expansion coefficients of the first pyramid block 21, the second pyramid block 22, and the third pyramid block 23 are 1, 2, and 4, respectively. The 3 pyramid blocks are used to capture three types of temporal features of the scattering center in the HRRP sequence: microscopic instantaneous fluctuations, mesoscopic structural evolution, and macroscopic attitude changes. The number of output channels of the 3 pyramid blocks are 64, 64, and 96, respectively.

[0029] This invention employs a pyramid multi-scale temporal convolutional network architecture, which captures three different scales of temporal features in the HRRP sequence: microscopic instantaneous fluctuations of the scattering center, mesoscopic structural evolution, and macroscopic attitude changes, through three cascaded pyramid blocks. This overcomes the shortcomings of traditional single-scale TCNs, which have a single receptive field and cannot simultaneously take into account local fine structure and global distribution patterns, and significantly improves the ability to distinguish between ship targets and corner reflector arrays.

[0030] See Figure 2 In each pyramid block, the main path is used to process the input features of the current pyramid block with two identical one-dimensional dilated convolutions, layer normalization, ReLU activation, and Dropout, with the Dropout rate ranging from 0.2 to 0.3, outputting the main path features. The secondary path is used to project the input features of the pyramid block to the same channel dimension as the output of the main path through a 1×1 convolutional layer, outputting the skip connection features. Each pyramid block adds the main path features and the skip connection features element-wise, and then performs ReLU activation and Dropout to obtain the output feature map of the current pyramid block.

[0031] See Figure 1 As one possible implementation, the multi-scale pooling fusion subunit 20 performs the following operations: Global average pooling is performed on the output feature map of each pyramid block along the temporal dimension to obtain N pooling vectors; By concatenating N pooling vectors along the channel dimension, a multi-scale fused feature vector is obtained.

[0032] The classification decision unit 3 is used to process the multi-scale fused feature vector and output the target category recognition result.

[0033] This invention employs a multi-scale feature fusion strategy combining global average pooling and channel concatenation. It integrates features from three pyramid blocks at different scales along the channel dimension, achieving complementary enhancement of multi-granularity information. Simultaneously, the pyramid blocks utilize a dual-path residual structure, effectively addressing the gradient vanishing problem in deep network training and ensuring training stability and feature reuse efficiency. The residual structure, through the parallel design of the main and secondary paths, allows gradients to propagate directly via skip connections, preventing degradation in deep networks. Global average pooling compresses features along the temporal dimension, preserving key information at each scale. Channel concatenation then organically fuses these multi-scale features, enabling the network to simultaneously utilize both microscopic details and macroscopic semantic information for comprehensive discrimination.

[0034] See Figure 1 As one possible implementation, the classification decision unit 3 includes a feature refinement network 30 and a classifier 31. The feature refinement network 30 includes two fully connected layers that map the multi-scale fused feature vectors into refined feature vectors. The classifier 31 maps the refined feature vectors to class scores and outputs the probability distribution of each class through the Softmax function. The dropout rate after the fully connected layer of the feature refinement network is 0.3, and the dropout rate after the fully connected layer of the classifier is 0.2.

[0035] As one possible implementation method, the target recognition results include three categories: ship targets, corner reflector arrays, and single corner reflectors.

[0036] See Figure 1 As an example, classifier 31 performs dimensionality reduction through two fully connected layers, each followed by a ReLU activation and a Dropout layer (Dropout rate 0.2), ultimately mapping to three categories and outputting the classification probability distribution through the Softmax function.

[0037] In a second aspect, embodiments of the present invention provide a method for identifying corner reflector arrays based on a pyramid multi-scale temporal convolutional network as described in the first aspect, comprising: S1. Obtain the raw data of the HRRP sequence to be identified; S2. Input the raw data of the HRRP sequence to be identified into the trained pyramid multi-scale temporal convolutional network; S3. The preprocessing unit performs amplitude normalization and initial feature extraction on the input HRRP sequence raw data, and outputs the initial feature map; S4. The feature extraction and fusion unit performs multi-scale feature extraction and pooling fusion on the initial feature map, and outputs a multi-scale fused feature vector; S5. The classification decision unit processes the multi-scale fused feature vector and outputs the target category recognition result, which includes three categories: ship target, corner reflector array, and single corner reflector.

[0038] As one possible implementation, before step S1, the pyramid multi-scale temporal convolutional network is trained. The training steps include: obtaining a training sample set, wherein the corner reflector array samples in the training sample set cover a variety of deployment scenarios, including but not limited to single array, multi-group collaboration, and random scattering, and the target pose angle covers 0°~180°; and training the pyramid multi-scale temporal convolutional network using a joint loss function that is a weighted combination of cross-entropy loss and center loss.

[0039] As one possible implementation, the joint loss function is as follows: in, This represents the equilibrium hyperparameter, with a value range of 0.01 to 0.2; The cross-entropy loss is calculated using the following method: In the formula, Indicates batch size, Indicates the total number of categories. Indicates the first The true label of each sample Indicates the first The sample belongs to the first The predicted probability of a class; The central loss is represented and calculated using the following method: In the formula, Indicates the first The refined feature vector of each sample Indicates the first Each sample corresponds to a feature center of its category.

[0040] This invention employs a joint loss function combining cross-entropy loss and center loss. While optimizing classification accuracy, the center loss constrains samples of the same class to cluster towards the class center in the feature space, while samples of different classes move away from each other. This effectively improves intra-class aggregation and inter-class separability in the feature space, enhancing the model's feature discrimination capability. Cross-entropy loss measures the difference between the predicted probability distribution and the true label, optimizing the classification boundary; center loss maintains the feature center of each class, constraining sample features to move closer to their corresponding class center during training. The joint optimization of these two loss functions enables the model to obtain more discriminative feature representations while improving classification accuracy, further reducing the confusion probability between ship targets and corner reflector arrays.

[0041] The technical solution and effects of the present invention will be further illustrated below with reference to specific embodiments.

[0042] Step 1: Construct a simulation sample set.

[0043] HRRP sequence data for typical ship targets, single corner reflectors, and corner reflector arrays were generated using electromagnetic simulation software; an example of the simulation dataset is shown below. Figure 3 As shown, the HRRP waveforms of a ship target, a single corner reflector, and a corner reflector array are displayed respectively. It can be seen that the HRRP waveform of the corner reflector array is highly similar to that of the ship target, making it difficult to effectively distinguish using traditional methods.

[0044] like Figure 4 As shown, the radial dimension of the ship target is 120 meters. The corner reflector array is configured with three deployment scenarios: single array, multiple coordinated arrays, and random scattering. The target attitude angle range is 0°~180° with a step size of 1°. 506 HRRP samples are generated for each deployment scenario. The simulation parameters are set as follows: carrier frequency 16GHz, bandwidth 64MHz, pulse width 7.2μs, and sampling rate 100MSPS.

[0045] Step 2: Normalize all HRRP samples, linearly mapping the amplitude of each sample to the [-1,1] interval. Use scaling, translation, and noise injection to augment the normalized samples, resulting in 5707 sets of sample data, including 1340 sets of ship samples, 1764 sets of single corner reflector samples, and 2604 sets of corner reflector array samples. Randomly divide the dataset into training, validation, and test sets in a ratio of 7:1.5:1.5.

[0046] Step 3: Train the pyramid multi-scale temporal convolutional network using a joint loss function that combines cross-entropy loss and center loss.

[0047] The joint loss function is as follows: in, This represents the equilibrium hyperparameter, with a value range of 0.01 to 0.2; The cross-entropy loss is calculated using the following method: In the formula, Indicates batch size, Indicates the total number of categories. Indicates the first The true label of each sample Indicates the first The sample belongs to the first The predicted probability of a class; The central loss is represented and calculated using the following method: In the formula, Indicates the first The refined feature vector of each sample Indicates the first Each sample corresponds to a feature center of its category.

[0048] Experimental platform: Windows 11 operating system, Intel Xeon Gold 6248R CPU, 128GB RAM, NVIDIA Tesla V100 GPU (32GB VRAM); Deep learning model implemented based on Python 3.10.

[0049] Training hyperparameter settings: The optimizer used is AdamW, and the initial learning rate is... Weight decay Batch size 32, total training epochs 200; learning rate using cosine annealing scheduling strategy, first 10 epochs as warm-up phase; global gradient pruning threshold set to 1.0; loss balancing hyperparameter .

[0050] Model training: Input the training set into the PM-TCN model, perform forward propagation to obtain classification prediction results, and calculate the joint loss function value; perform backpropagation through the AdamW optimizer to update the network parameters; after each round of training, verify the model performance through the validation set and save the model weights with the best performance on the validation set; repeat training until 200 rounds of training are completed.

[0051] Step 4: Use the trained pyramid multi-scale temporal convolutional network to perform initial shallow feature extraction on the HRRP sequence.

[0052] Dimension is HRRP sequence ( The input is a one-dimensional convolutional layer (where the number of distance gates is 3), with a kernel size of 3, using same padding to preserve sequence length, and an output channel count of 32. After passing through layer normalization and the ReLU activation function, the resulting layer has a dimension of... initial feature map ,in, For batch size, The sequence length is given.

[0053] Step 5: Multi-scale feature extraction and pooling fusion to obtain multi-scale fused feature vectors.

[0054] The three pyramid blocks have expansion coefficients of 1, 2, and 4, respectively, a kernel size of 3, same padding, and output channels of 64, 64, and 96, respectively. The processing flow for each pyramid block is as follows: Let the first The input for the pyramid block is , , This indicates the number of output channels of the preprocessing unit.

[0055] Main path: First, it passes through a one-dimensional dilated convolutional layer with an inflation coefficient of [missing information]. Kernel size 3, same padding, number of output channels The convolutional, normalization, ReLU activation, and Dropout layers (with a Dropout rate of 0.2) are then applied sequentially. This process is repeated once more to obtain the main path output. ; Secondary path: via The convolutional layer will input Project to With the same channel dimension, skip connection features are obtained. ; Will and Element-wise addition is performed, followed by ReLU activation and a Dropout layer to obtain the output feature map of the current pyramid block. .

[0056] Feature maps output from the three pyramid blocks , , Along the time-series dimension Perform global average pooling to obtain the pooling vector. , , The three pooling vectors are concatenated along the channel dimension to obtain a multi-scale fused feature vector. .

[0057] Step 6: The classification decision unit processes the multi-scale fused feature vector and outputs the target category recognition result.

[0058] The feature refinement network consists of two fully connected layers: the first fully connected layer maps the 224-dimensional features to 256 dimensions, followed by ReLU activation and a Dropout layer (Dropout rate 0.3); the second fully connected layer maps the 256-dimensional features to 128 dimensions, followed by ReLU activation and a Dropout layer (Dropout rate 0.3), resulting in the refined feature vector. The classifier consists of two fully connected layers: the first layer reduces the 128-dimensional features to 64 dimensions, followed by ReLU activation and a Dropout layer; the second layer reduces the 64-dimensional features to 32 dimensions, followed by ReLU activation; finally, the 32-dimensional features are mapped to scores for the three categories. The category scores are converted into a probability distribution using the Softmax function. The category corresponding to the highest probability is selected as the final recognition result, completing the classification and recognition of ship targets, corner reflector arrays, and single corner reflectors.

[0059] The experimental results of this invention (PM-TCN) are compared with those of models such as LSTM, SSM, and ResNet18. Figure 5 As shown, PM-TCN outperforms comparable models such as LSTM, SSM, and ResNet18 under all signal-to-noise ratio (SNR) conditions, with a particularly significant advantage under low SNR conditions of 0dB to 4dB. The overall recognition accuracy is improved by at least 10 percentage points compared to the comparative algorithms. The confusion matrix of the PM-TCN model at a SNR of 20dB is shown below. Figure 6 As shown in the figure. The results show that the model achieves recognition accuracy of 96.7%, 97.5%, and 98.7% for ships, corner reflector arrays, and single corner reflectors, respectively, with an overall F1-score of 0.9749 and extremely low confusion rates between categories, verifying the model's excellent recognition performance.

[0060] It should be noted that when When the value is taken in the range of 0.01 to 0.2, the joint loss function of this invention can effectively balance classification accuracy and feature discriminativeness, achieving a recognition effect superior to that of a single loss function. The optimal overall performance is shown in Table 1: Table 1 Differences The influence of the value on the performance of the invention To analyze the contribution of features at different scales (values ​​of N) to recognition performance, this invention sets up multiple sets of comparison structures: single-branch pyramid blocks (N=1), dual-branch (N=2), triple-branch (PM-TCN, N=3), and quadruple-branch (N=4) combinations. By comparing the performance of each combination on the same test set (signal-to-noise ratio of 16dB), the necessity of multi-scale feature joint analysis is verified, and the results are shown in Table 2.

[0061] As shown in Table 2, the single-branch structure exhibits the weakest performance, indicating that relying solely on a single-scale receptive field is insufficient to comprehensively characterize the differences between ship targets and corner reflector arrays. From single-branch to four-branch, the combined performance continuously improves with the increase in the number of branches, demonstrating that introducing more scale information helps enhance feature representation capabilities. PM-TCN (N=3) achieves the best performance-efficiency tradeoff, with its F1-score (0.9749) almost matching that of the four-branch TCN (0.9770), but with a 23% reduction in inference time, making it more valuable for applications.

[0062] Table 2. Effects of different values ​​of N on the performance of the invention. Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the description of the drawings, in carrying out the claimed invention. In this specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple components. A single processor or other unit can implement several of the functions listed in the specification. While certain measures are described in different embodiments, this does not mean that these measures cannot be combined to produce good results.

[0063] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely illustrative of the invention and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A pyramid-shaped multi-scale temporal convolutional network, characterized in that, It includes a preprocessing unit, a feature extraction and fusion unit, and a classification decision unit connected sequentially. The preprocessing unit is used to perform amplitude normalization and initial feature extraction on the input HRRP sequence raw data, and output the initial feature map; The feature extraction and fusion unit comprises N cascaded pyramid blocks and a multi-scale pooling fusion subunit, where N ≥ 2. The N pyramid blocks employ progressively increasing expansion coefficients. Each pyramid block includes a parallel main path and secondary paths. The main path is used to extract temporal features at the corresponding scale of the pyramid block, while the secondary paths are used to perform dimensional mapping on the input features to make them consistent with the output dimension of the main path. The multi-scale pooling fusion subunit is used to perform multi-scale fusion on the output feature maps of each pyramid block to obtain a multi-scale fused feature vector. The classification decision unit is used to process the multi-scale fused feature vector and output the target category recognition result.

2. The pyramid-shaped multi-scale temporal convolutional network according to claim 1, characterized in that, N = 3, with expansion coefficients of 1, 2 and 4 for the three pyramid blocks, respectively, used to capture three types of temporal features of the scattering center in the HRRP sequence: microscopic instantaneous fluctuations, mesoscopic structural evolution and macroscopic attitude changes. The number of output channels for the three pyramid blocks are 64, 64 and 96, respectively.

3. The pyramid multi-scale temporal convolutional network according to claim 2, characterized in that, In each pyramid block, the main path is used to process the input features of the current pyramid block with two identical one-dimensional dilated convolutions, layer normalization, ReLU activation and Dropout, with the Dropout rate ranging from 0.2 to 0.3, and outputs the main path features; the secondary path is used to project the input features of the pyramid block to the same channel dimension as the output of the main path through a 1×1 convolutional layer, and outputs the skip connection features. Each pyramid block adds the main path features and skip connection features element-wise, then performs ReLU activation and Dropout to obtain the output feature map of the current pyramid block.

4. The pyramid multi-scale temporal convolutional network according to claim 1, characterized in that, The multi-scale pooling fusion subunit performs the following operations: Global average pooling is performed on the output feature map of each pyramid block along the temporal dimension to obtain N pooling vectors; By concatenating N pooling vectors along the channel dimension, a multi-scale fused feature vector is obtained.

5. The pyramid multi-scale temporal convolutional network according to claim 1, characterized in that, The preprocessing unit includes a one-dimensional convolutional layer, a layer normalization layer, and a ReLU activation layer. The one-dimensional convolutional layer has 32 output channels, and the convolutional kernel uses same padding to keep the sequence length unchanged. The normalization process linearly maps the amplitude of each sample in the HRRP sequence to the interval [-1, 1].

6. The pyramid-shaped multi-scale temporal convolutional network according to claim 1, characterized in that, The classification decision unit includes a feature refinement network and a classifier. The feature refinement network includes two fully connected layers that map the multi-scale fused feature vectors into refined feature vectors. The classifier maps the refined feature vectors to class scores and outputs the probability distribution of each class through a Softmax function. The dropout rate after the fully connected layer of the feature refinement network is 0.3, and the dropout rate after the fully connected layer of the classifier is 0.

2.

7. The pyramid multi-scale temporal convolutional network according to claim 1, characterized in that, The target recognition results include three categories: ship targets, corner reflector arrays, and single corner reflectors.

8. A method for identifying corner reflector arrays based on the pyramid multi-scale temporal convolutional network according to any one of claims 1 to 7, characterized in that, include: S1. Obtain the raw data of the HRRP sequence to be identified; S2. Input the raw data of the HRRP sequence to be identified into the trained pyramid multi-scale temporal convolutional network; S3. The preprocessing unit performs amplitude normalization and initial feature extraction on the input HRRP sequence raw data, and outputs the initial feature map; S4. The feature extraction and fusion unit performs multi-scale feature extraction and pooling fusion on the initial feature map, and outputs a multi-scale fused feature vector; S5. The classification decision unit processes the multi-scale fused feature vector and outputs the target category recognition result, which includes three categories: ship target, corner reflector array, and single corner reflector.

9. The corner reflector array identification method according to claim 8, characterized in that, Before step S1, the training of the pyramid multi-scale temporal convolutional network is also included. The training steps include: obtaining a training sample set, wherein the corner reflector array samples in the training sample set cover a variety of deployment scenarios, including but not limited to single array, multi-group collaboration, and random scattering, and the target pose angle covers 0°~180°; and training the pyramid multi-scale temporal convolutional network using a joint loss function that is a weighted combination of cross-entropy loss and center loss.

10. The corner reflector array identification method according to claim 9, characterized in that, The joint loss function is as follows: in, This represents the equilibrium hyperparameter, with a value range of 0.01 to 0.2; The cross-entropy loss is calculated using the following method: In the formula, Indicates batch size, Indicates the total number of categories. Indicates the first The true label of each sample Indicates the first The sample belongs to the first The predicted probability of a class; The central loss is represented and calculated using the following method: In the formula, Indicates the first The refined feature vector of each sample Indicates the first Each sample corresponds to a feature center of its category.

Images