Small sample radar target recognition method based on multi-feature fusion

Abstract

The present invention discloses a small-sample radar target recognition method based on multi-feature fusion. The present invention introduces the natural resonant frequency characteristics of the radar echo signal into the small-sample learning framework, making full use of its advantages of certain stability under the conditions of attitude and observation angle changes, and simple and fast extraction; and performs energy-guided attention multi-feature weighted fusion of the natural resonant frequency characteristics and time-frequency characteristics, describing the scattering characteristics from different feature mechanisms, and improving the completeness of the features under small sample conditions; in addition, a loss function based on the composite of feature similarity measurement and classification loss is designed, which optimizes the discriminability of the feature space while facilitating the model to learn more discriminative feature representations, thereby improving performance in the feature fusion-classification network. The present invention effectively improves the recognition accuracy under the condition of observation angle changes without significantly increasing the inference time and computational complexity.

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal processing and radar target recognition, and in particular to a small-sample radar target recognition method based on multi-feature fusion. Background Art

[0002] Radar target recognition technology, with its advantages of all-weather operation, strong penetration, and high-resolution feature extraction, holds broad application prospects in defense, military, remote sensing, and other fields. With the development of fields like artificial intelligence and cognitive radar, its accuracy and intelligence are continuously improving. Currently, the big data-based supervised learning paradigm relies heavily on the scale and quality of labeled data, severely restricting the effectiveness of radar target recognition technology in scenarios where labeled resources are limited. For applications such as disaster monitoring, geological exploration, early warning detection, and camouflaged target resolution, radar primarily identifies non-cooperative targets. This is compounded by long data collection and processing cycles, unstable and inaccurate feature extraction, and high annotation costs and expertise requirements. These limitations severely limit traditional deep learning approaches leveraging big data. Against this backdrop, exploiting the advantages of few-shot learning (FSL) has become a pressing need for radar target recognition applications.

[0003] Currently, mainstream small-shot learning approaches in radar target recognition primarily utilize features such as Synthetic Aperture Radar (SAR) images and High Resolution Range Profile (HRRP). However, these features are complex to extract and computationally inefficient. Furthermore, they lack stability when the observation angle or target pose changes, making them unable to overcome the dramatic feature fluctuations in complex electromagnetic environments. This makes it difficult to meet the demand for small-shot learning models to extract discriminative and stable features from limited samples. Furthermore, existing small-shot learning methods for radar target recognition often rely on a single feature or fuse multiple statistical features of the same feature data. This is essentially the fusion of different features operating under the same mechanism, often exhibiting a certain degree of feature correlation. This makes it difficult to establish reasonable feature associations, and the resulting fused features lack sufficient completeness. Furthermore, existing multi-feature fusion-classification methods lack sufficient complementarity between metric space optimization and loss function design in terms of spatial discriminative constraints, often struggling to balance feature discriminability enhancement with the requirements of the classification task. Summary of the Invention

[0004] In view of this, the present invention provides a small sample radar target recognition method based on multi-feature fusion, which takes the natural resonant frequency feature of the resonance zone as the core recognition feature and fuses it with the signal time-frequency feature. The fused feature is stable and complete, and can effectively realize the recognition of small sample radar targets.

[0005] The small sample radar target recognition method based on multi-feature fusion of the present invention includes:

[0006] S1, extract the time-frequency image and natural resonance frequency of the radar echo signal;

[0007] S2, constructing a multi-feature fusion classification network; the multi-feature fusion classification network includes a multi-layer convolutional neural network, a recurrent neural network, a feature fusion network and a classifier;

[0008] Among them, the multi-layer convolutional neural network is used as an encoder to extract the time-frequency features of the time-frequency image;

[0009] Recurrent neural network is used as an encoder to extract natural resonant frequency features;

[0010] The feature fusion network concatenates the time-frequency features and the natural resonant frequency features and multiplies them by a weight matrix to generate fused features. The weight submatrix of the time-frequency features is 1, while the weight submatrix of the natural resonant frequency features is: The modal energy and total energy of each frequency in the natural resonant frequency features are calculated, and then energy-guided modal saliency weights are generated based on a differentiable attention mechanism. The weights are normalized to obtain the weight submatrix of the natural resonant frequency features.

[0011] The classifier classifies the target based on the fused features;

[0012] S3, trains the multi-feature fusion classification network constructed in S2; the loss function is the weighted sum of triple loss, dynamic time warping loss and cross entropy loss;

[0013] S4, uses the trained multi-feature fusion classification network to complete target recognition.

[0014] Preferably, in S1, the time-frequency image of the radar echo signal is obtained by using the Cui-Williams distribution time-frequency analysis, the Wigner-Wiley distribution time-frequency analysis or the short-time Fourier transform.

[0015] Preferably, in S1, the time domain / frequency domain Prony method, matrix bundle method, iterative method or Cauchy method is used to extract the natural resonant frequency of the radar echo signal.

[0016] Preferably, in S2, the multi-layer convolutional neural network uses CNN, ResNet-18 or ResNet-50 to extract the time-frequency features of the time-frequency image.

[0017] Preferably, in S2, the recurrent neural network uses RNN, LSTM or BiLSTM to extract the natural resonance frequency features.

[0018] Preferably, the feature fusion network concatenates the time-frequency features and the natural resonant frequency features, and then multiplies them element-by-element with the weight matrix to generate fused features.

[0019] Preferably, in S2, the weight coefficient of the triplet loss is not less than the weight coefficient of the dynamic time warping loss; the weight coefficient of the cross entropy loss is not less than the sum of the weight coefficients of the triplet loss and the dynamic time warping loss.

[0020] Preferably, the weight coefficient ratios of dynamic time warping loss, triplet loss, and cross entropy loss are 1:2:7, 1:1:8, 2:2:6, or 2:3:5.

[0021] Preferably, in S3, triplet loss and DTW dynamic time warping loss use Wasserstein distance, Euclidean distance, Manhattan distance, Chebyshev distance or cosine similarity to measure feature similarity.

[0022] Preferably, during S3 model training, use the SGD optimizer, Adam optimizer, or RMSProp optimizer.

[0023] Beneficial effects:

[0024] In response to the limitations of traditional meta-learning methods that over-focus on model generalization ability while ignoring the ability to express features, the present invention introduces the natural resonance frequency sequence that characterizes the scattering characteristics of the target resonance area into the small-sample learning framework, fully utilizing its advantages of certain stability under conditions of posture and observation angle changes, as well as simple and fast extraction. The natural resonance frequency features are fused with time-frequency features to overcome the negative impact of the natural resonance frequency extraction accuracy, and the scattering features are described from different feature mechanisms to improve the completeness of features under small-sample conditions. At the same time, the energy of each natural resonance frequency mode is used as a weight coefficient matrix to weight the natural resonance frequency features through the Energy-Guided Attentive Fusion (EGAF) mechanism, thereby achieving a deep fusion of time-frequency features and natural resonance frequency features to achieve the final classification task. In addition, a CFE Loss function based on a combination of feature similarity measurement and classification loss is designed. While optimizing the discriminability of the feature space, it is beneficial for the model to learn more discriminative feature representations, thereby improving performance in the feature fusion-classification network.

[0025] Experimental results demonstrate that, without significantly increasing inference time and computational complexity, the proposed method achieves a recognition accuracy of 95.36% under varying observation angles, representing improvements of 1.66% and 8.73% over traditional single-feature and concatenated feature fusion methods, respectively. Ablation experiments further validate the effectiveness of each innovative module, demonstrating the feasibility and practical value of the proposed method. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 Flow chart of the method of the present invention.

[0027] Figure 2 Schematic diagram of the time-frequency feature extraction structure based on ResNet-18.

[0028] Figure 3 Schematic diagram of the natural resonance frequency feature extraction structure based on BiLSTM.

[0029] Figure 4 Schematic diagram of the fusion structure of natural resonant frequency features and time-frequency features based on energy-guided attention weights. DETAILED DESCRIPTION

[0030] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0031] The present invention provides a small sample radar target recognition method based on multi-feature fusion.

[0032] The present invention uses the natural resonant frequency characteristics of the radar echo signal as the core recognition feature. This feature reflects the intrinsic electromagnetic scattering characteristics of the target in the resonant region and can characterize the target's resonant scattering characteristics in an extremely low dimension. Its inherent physical properties and scattering mechanism give it a certain degree of invariance in posture and observation angle, and it has stability, which is more in line with the generalization requirements of small sample learning for features. At the same time, the natural resonant frequency characteristics are fused with the time-frequency characteristics of the radar signal for multi-feature recognition. These two features have good distinguishability, can achieve multi-dimensional decoupling and reconstruction of the target scattering characteristics, and have good completeness, making it better suitable for the effective identification of small sample radar targets in complex environments.

[0033] Specifically, the method of the present invention is shown in the flowchart 1, which includes the following steps:

[0034] Step 1: extract the time-frequency image and natural resonant frequency of the radar echo signal;

[0035] Perform time-frequency analysis on the radar echo signal, such as Choi-Williams Distribution (CWD) time-frequency analysis, Wigner-Ville Distribution (WVD) time-frequency analysis, and Short-Time Fourier Transform (STFT), to obtain a time-frequency image.

[0036] At the same time, the matrix beam method, time domain / frequency domain Prony method, iterative method, Cauchy method, etc. are used to extract the natural resonant frequency of the radar echo signal.

[0037] Step 2: Build a multi-feature fusion classification network;

[0038] The present invention constructs a collaborative coding process based on time-frequency characteristics and natural resonant frequency characteristics, and jointly describes the scattering characteristics of the target from two different dimensions: the energy distribution of the echo signal and the resonant oscillation change, thereby realizing multi-dimensional decoupling and reconstruction of the target scattering characteristics.

[0039] Firstly, a multi-layer convolutional neural network is used to extract the energy aggregation features in the frequency domain layer by layer using a multi-layer convolution kernel group to obtain the time-frequency features of the time-frequency image; and a recurrent neural network is used to capture the oscillation change law and angular stability characteristics of the resonance pole over time to obtain the natural resonance frequency characteristics.

[0040] Multi-layer convolutional neural networks can use CNN networks, residual neural networks, etc., with Residual Neural Networks (ResNet) preferably using ResNet-18 and ResNet-50. Recurrent neural networks can use RNN, LSTM, BiLSTM, etc.

[0041] Then, a multi-feature fusion network is used to splice the time-frequency features and the natural resonant frequency features, and EGAF is used to enhance the characteristic response of the key frequency band of the natural resonant frequency, thereby enhancing the model's ability to characterize the physical nature of the signal.

[0042] With training samples As an example, let its time-frequency image features and natural resonance frequency features be and , the multi-feature fusion network splices the two features together to construct the features of the sample:

[0043] (1)

[0044] in, Indicates time-frequency analysis, and the local feature dimension is 256. Among them, the time-frequency image features of the sample The natural resonance frequency features are extracted by multi-layer convolutional neural network. Extracted by recurrent neural network.

[0045] To more comprehensively describe the signal's physical characteristics and enhance the model's ability to focus on learning fused features, the fused features are weighted based on signal energy. Time-frequency features are extracted based on the signal's energy distribution, so their weights are assigned a value of 1. For natural resonant frequency features, the modal energy and total energy of each frequency in the natural resonant frequency feature are first calculated, and then the weight of that frequency feature is calculated based on the proportion of that frequency's modal energy in the total energy.

[0046] Specifically, according to the singular point expansion theory, the time response of the radar target resonance region can be expressed by the sum of the attenuation negative exponentials of the natural resonance frequency. The natural resonant frequency, its mode is ,in, For the remainder, is the natural resonant frequency, is the attenuation factor. Then its energy for:

[0047] (2)

[0048] Due to the stability requirements of the system , the above integral converges, so the energy can be further simplified to

[0049] (3)

[0050] This result shows that the energy of each natural resonant frequency mode is proportional to the square of its amplitude and inversely proportional to the absolute value of the decay rate. In this way, the energy of each natural resonant frequency mode can be used as a weight matrix for subsequent signal analysis and feature extraction. By summing the energies of all natural resonant frequency modes, a global energy scalar or matrix is ​​obtained E , which is used for subsequent weight distribution. The total energy of the signal can be obtained by summing the energies of all natural resonant frequency modes:

[0051] (4)

[0052] Generating energy-guided modality saliency weights via a differentiable attention mechanism μ , normalize the relevance scores to a probability distribution so that the weights μ satisfy , the calculation method is:

[0053] (5)

[0054] The weight embeds the energy information of each mode of the natural resonant frequency into the fusion feature. Modes with larger amplitudes and slower decay rates occupy more energy in the signal, so they are given higher weights in the weight matrix. The matrix can intuitively reflect the relative importance of different natural resonant frequency modes in the signal. Highlighting the features of important natural resonant frequency modes and suppressing redundant information will help subsequent models capture and learn the key features of the signal.

[0055] The final weight matrix The dimension is 512:

[0056] (6)

[0057] Modulated fusion features for:

[0058] (7)

[0059] Here, ⊙ represents element-wise multiplication (Hadamard product). This product method utilizes a broadcast mechanism for modality-by-modality scaling, which can reduce redundant information in a noise-free environment. Finally, a classifier is used to map the fused features into a category space through a fully connected layer, outputting a score for each category.

[0060] Step three: train the constructed multi-feature fusion classification network.

[0061] During training, a Classification-Fusion Equilibrium Loss (CFE Loss) function is designed, combining a feature similarity metric with a classification loss. The CFE Loss is a weighted combination of the triplet loss, dynamic time warping (DTW), and cross-entropy loss. The triplet loss and DTW loss, respectively, serve as similarity metrics for time-frequency features (image features) and natural resonant frequency features (sequence features) to enhance feature discriminability, while the cross-entropy loss optimizes the model's classification performance. The DTW loss enhances the stability of the natural resonant frequency through elastic temporal alignment, complementing the spatial discriminative constraint of the triplet loss. The similarity metric can be viewed as an implicit regularization term, suppressing overfitting of the model to a single feature. By constraining the spatial similarity of multiple features through the similarity metric, the model is encouraged to learn shared representations of different features, enhancing the complementarity of information between them. At the same time, the cross entropy loss ensures that the model has a certain degree of discriminability in the classification task, preventing the similarity optimization from causing deviation from the target task.

[0062] The design of the weight coefficients can avoid loss function optimization direction deviation and place the optimization focus on the classification task. To achieve the coordinated optimization of classification loss and feature regularization, the initial focus is on feature distribution modeling, and the weight coefficient of the triplet loss is not less than the weight coefficient of the dynamic time warping loss. Later, the classification boundary optimization is strengthened, with the classification task as the primary goal, and the weight coefficient of the cross-entropy loss is not less than the sum of the weight coefficients of the triplet loss and the dynamic time warping loss. The preferred weight coefficient ratios of the dynamic time warping loss, triplet loss, and cross-entropy loss are 1:2:7, 1:1:8, 2:2:6, and 2:3:5.

[0063] Triplet loss and DTW loss can use Wasserstein distance, Euclidean distance, Manhattan distance, Chebyshev distance, cosine similarity and other metrics to optimize the difference in feature distribution. Even when the two feature distributions are very different or do not completely overlap, effective gradient information can be calculated, which has a positive effect on maintaining stability during model optimization.

[0064] The preferred optimizers are SGD, Adam, and RMSProp.

[0065] Step 4: Complete the classification and recognition of the target based on the trained multi-feature fusion classification network.

[0066] Example

[0067] The following example uses a simulated data set to perform radar target recognition under small sample conditions to verify the effectiveness of the method proposed in the present invention. The method specifically includes the following steps:

[0068] Step 1: Use FEKO software to perform target simulation and data acquisition.

[0069] A number of simple models were simulated using FEKO, including nine types of targets such as spheres, cones, rods, and rectangular plates, all made of copper. Under far-field, single-station conditions, the frequency range was set to 1MHz-2.7GHz, the frequency point was set to sweep frequency, the observation angle interval was set to 5°, and the angle variation range was The frequency domain signal of the target is obtained by calculating the Method of Moments (MoM), and the corresponding time domain echo is obtained by inverse Fourier transform.

[0070] Step 2: Extract the time-frequency image and natural resonant frequency of the echo.

[0071] The echo is subjected to Cui-Williams distribution time-frequency analysis to obtain a time-frequency image.

[0072] At the same time, the joint matrix bundle method is used to extract the natural resonance frequency of the echo, as follows:

[0073] According to the singular point expansion theory, the time-lapse response of a radar target in the resonance region can be expressed as the sum of the decaying complex exponentials of the extreme points, where the transfer function of the target is expressed as:

[0074] (8)

[0075] in, is the number of extreme points, and denote the extreme parts of the residue and the target respectively, Defined as the attenuation factor. The time domain signal is obtained by windowing the obtained frequency domain signal and performing inverse Fourier transform , construct the Hankel matrix :

[0076] (9)

[0077] in, represents the observation angle, is the matrix bundle parameter, here we take Arrange the Hankel matrices obtained at each angle into a multi-angle matrix , and perform a singular value decomposition on it:

[0078] (10)

[0079] (11)

[0080] According to the above formula, construct 、 、 matrix:

[0081] (12)

[0082] (13)

[0083] (14)

[0084] Constructing a Matrix :

[0085] (15)

[0086] (16)

[0087] By seeking The generalized eigenvalues ​​of the joint matrix are used to obtain the poles of the joint matrix, which are the natural resonant frequencies of the radar echo signal. Poles are insensitive to angle, and while the joint matrix under multi-angle observation theoretically has corresponding poles, it is difficult to avoid the presence of noise with large singular values, which can interfere with the extraction process and affect the results as false poles. Since this research does not focus on pole extraction, false poles are not removed in the subsequent identification process.

[0088] The generated dataset contains 2019 samples of nine categories of targets, and the training set and test set are divided into 8:2. The model randomly selects 1615 small sample tasks from the dataset as the training set, and the remaining 404 samples form the test set to evaluate the model performance.

[0089] Step 3: Build a Multi-feature Fusion Classification Network based on BiLSTM and ResNet (Res-BiLSTM-MFCNet) based on the residual neural network ResNet-18 and the bidirectional long short-term memory neural network (BiLSTM). Figure 1 shown.

[0090] like Figure 2 As shown in the figure, the convolution kernel group of the residual neural network ResNet-18 extracts the time-frequency features of the time-frequency image layer by layer. The processing flow is as follows:

[0091] Input single-channel time-frequency grayscale image ;

[0092] The initial convolution layer uses a large receptive field convolution kernel to extract low-frequency energy distribution features. The convolution kernel configuration is , kernel=7 7, stride=2, padding=3, output feature map After normalization and activation ; After a pooling layer MaxPool2d, kernel=3 3, stride=2, padding=1, output dimension The residual network layer group realizes feature abstraction through a four-level residual structure, where Layer 1 consists of two standard residual blocks with the structure , the output dimension is 64×64×64. The remaining three layers each contain two downsampled Bottleneck layers, which are 1 Dimensionality Reduction-3 3 convolution-1 1 Dimensionality reduction: The residual block uses stride=2 convolution to match the dimension, and the output dimension sequences are 128×32×32, 256×16×16, and 512×8×8. Finally, a global pooling and fully connected layer are connected to perform feature compression and dimensionality reduction, and the image feature is 256-dimensional.

[0093] like Figure 3 As shown, BiLSTM passes through the forward and backward hidden states , To capture the contextual dependencies of the natural resonant frequency sequence, the processing flow is as follows:

[0094] Input natural resonant frequency imaginary part sequence ;

[0095] The time series embedding layer performs linear projection, resulting in an output dimension of 64. Each hidden unit in the two-layer BiLSTM is 128-dimensional. The embedding layer maps the scalar sequence into a high-dimensional space, enhancing its nonlinear representation capabilities. Through forward and backward propagation, the BiLSTM can simultaneously represent the forward resonance buildup and backward resonance decay processes. Features are concatenated to output a 256-dimensional feature vector.

[0096] The structural diagram of the multi-feature fusion network is as follows Figure 4 As shown in the figure, based on the energy-guided attention EGAF framework, the amplitude and attenuation rate of the natural resonant frequency are converted into a fusion feature weight matrix representing the energy, the natural resonant frequency features are weighted, and the energy-attention-weighted natural resonant frequency features and time-frequency features are spliced ​​to obtain the fusion features.

[0097] Step 5: Construct training sample pairs based on the support set and query set, and use CFE Loss as the loss function.

[0098] (17)

[0099] in, is the weight coefficient; is the dynamic time warping loss; is the triplet loss; is the cross entropy loss. In this embodiment, =1:2:7.

[0100] Use the more common 5-way 1-shot The small sample classification sampling strategy randomly selects 6 samples from each of the five categories in the training set, meaning each category has 5 support set samples and 1 query set sample. Training is performed end-to-end using the SGD optimization algorithm with a batch size of 32, a learning rate of 0.001, and a momentum of 0.9 for 100 iterations, with all conditions remaining constant.

[0101] Selecting Wasserstein distance as the distance metric for time-frequency features / natural resonant frequency features can measure and optimize the difference in feature distributions. Even when the two feature distributions are quite different or not completely overlapping, effective gradient information can be calculated, which has a positive effect on maintaining stability during model optimization. The characteristic probability distributions of When , the Wasserstein distance metric calculation formula is:

[0102] (18)

[0103] Where, express The joint distribution of represents the set of all possible joint distributions, For respectively Randomly selected variables, is the infimum, which means finding the minimum expected distance.

[0104] Wasserstein distance can measure and optimize the difference between the distributions of two types of features. Even when the two feature distributions differ greatly or do not completely overlap, effective gradient information can be calculated, which has a positive effect on maintaining stability during model optimization. Therefore, the specific calculation formula for triple loss based on Wasserstein distance is as follows:

[0105] (19),

[0106] in, is the feature representation of the anchor sample, is the feature representation of the positive sample of the same type as the anchor point, is the negative sample feature representation that is different from the anchor point, is the margin, which specifies the minimum required separation between positive and negative sample pairs. is the distance metric calculation of formula (18). Indicates that the loss is non-negative and is 0 when the model meets the expected conditions. Otherwise, the loss is greater than 0, which means that the model is penalized. The optimization goal of this loss function is to maximize the Wasserstein distance between the anchor point and the negative sample and minimize the Wasserstein distance between the anchor point and the positive sample.

[0107] The experimental verification results are shown in Table 1.

[0108] Table 1 Comparative experimental results of different features

[0109]

[0110] As can be seen from Table 1, the fusion features used in the present invention perform best in terms of recognition accuracy and stability. Although its relatively complex model structure leads to an increase in training and inference time, it is still the best choice in scenarios with high accuracy requirements. When a single time-frequency feature is used as input, the accuracy and inference time are relatively balanced, and the inference time is reduced by 17.3%. Although the single natural resonance frequency feature model has the advantages of small data volume and fast calculation, it performs poorly under small sample learning. This is mainly because the feature distribution of the natural resonance frequency data does not have a significant clustering effect, the inter-class differences are small, and the model is difficult to learn the physical properties contained therein. Without going through the encoding-decoding process, the two features are directly tensor spliced ​​to generate a spliced ​​fusion feature. The splicing fusion model using this feature performs generally in all indicators. The method of directly splicing tensors for two-dimensional images and sequence data is simple and intuitive, but image data and structured sequence data have different feature spaces and semantic information. The model is difficult to capture the correlation between the two features, and the features after direct splicing also lack clear physical meaning, resulting in low recognition accuracy. From the comparison of parameter quantity indicators, the features used in the present invention are only slightly higher than those using single time-frequency features and spliced ​​fusion features, while the FLOPS are the same, proving that the resource consumption and computational complexity of the model of the present invention have not increased significantly.

[0111] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A small sample radar target recognition method based on multi-feature fusion, characterized in that: include: S1, extract the time-frequency image and natural resonance frequency of the radar echo signal; S2, constructing a multi-feature fusion classification network; the multi-feature fusion classification network includes a multi-layer convolutional neural network, a recurrent neural network, a feature fusion network and a classifier; Among them, the multi-layer convolutional neural network is used to extract the time-frequency features of the time-frequency image; Recurrent neural network is used to extract natural resonant frequency features; The feature fusion network concatenates the time-frequency features and the natural resonant frequency features and multiplies them by the weight matrix to generate fusion features; the weight sub-matrix of the time-frequency features is 1; the weight sub-matrix of the natural resonant frequency features is μ for: ， in, is the energy of the kth natural resonant frequency mode, ,in, For the remainder, is the attenuation factor; is the energy sum of all natural resonant frequency modes; The classifier classifies the target based on the fused features; S3, trains the multi-feature fusion classification network constructed in S2; the loss function is the weighted sum of triple loss, dynamic time warping loss and cross entropy loss; S4, uses the trained multi-feature fusion classification network to complete target recognition.

2. The method according to claim 1, wherein In S1, the time-frequency image of the radar echo signal is obtained by using the Cui-Williams distribution time-frequency analysis, the Wigner-Wiley distribution time-frequency analysis or the short-time Fourier transform.

3. The method according to claim 1, wherein In S1, the time domain / frequency domain Prony method, matrix bundle method, iterative method or Cauchy method is used to extract the natural resonant frequency of the radar echo signal.

4. The method according to claim 1 or 2, wherein: In S2, the multi-layer convolutional neural network uses CNN, ResNet-18 or ResNet-50 to extract the time-frequency features of the time-frequency image.

5. The method according to claim 1 or 3, wherein In S2, the recurrent neural network uses RNN, LSTM or BiLSTM to extract natural resonance frequency features.

6. The method according to claim 1, wherein The feature fusion network concatenates the time-frequency features and the natural resonant frequency features, and then multiplies them element-by-element with the weight matrix to generate fused features.

7. The method according to claim 1, wherein In S2, the weight coefficient of the triplet loss is not less than the weight coefficient of the dynamic time warping loss; the weight coefficient of the cross entropy loss is not less than the sum of the weight coefficients of the triplet loss and the dynamic time warping loss.

8. The method according to claim 7, wherein The weight coefficient ratios of dynamic time warping loss, triplet loss, and cross entropy loss are 1:2:7, 1:1:8, 2:2:6, or 2:3:

5.

9. The method according to claim 1, 7 or 8, wherein: In S3, triplet loss and DTW dynamic time warping loss use Wasserstein distance, Euclidean distance, Manhattan distance, Chebyshev distance or cosine similarity to measure feature similarity.

10. The method according to claim 1, wherein During S3 model training, the SGD optimizer, Adam optimizer, or RMSProp optimizer is used.

Images