Echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraints

Abstract

This invention discloses an echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraints, comprising: acquiring radar multimodal echo data of a target to be identified; preprocessing the radar multimodal echo data to obtain first radar multimodal echo data; inputting the first radar multimodal echo data into a trained adaptive wavelet denoising and identification integrated model to obtain the target identification result corresponding to the radar multimodal echo data; wherein, the adaptive wavelet denoising and identification integrated model is used to perform signal structure decomposition and denoising on the input first radar multimodal echo data to enhance the echo, and to fuse and classify the enhanced echo features. The method of this application solves the problems of radar echo feature distortion and recognition performance degradation under low signal-to-noise ratio conditions, achieves synergistic optimization of noise suppression and target identification, and improves the accuracy and robustness of radar target identification in complex electromagnetic environments.

Description

Technical Field

[0001] This invention relates to the field of radar technology, and in particular to a method for echo enhancement and robust identification with low signal-to-noise ratio based on signal structure constraints. Background Technology

[0002] Radar target recognition technology has significant applications in military reconnaissance, air traffic control, and meteorological monitoring. Its performance largely depends on the quality of the training data, typically requiring the use of high signal-to-noise ratio radar one-dimensional high-resolution range profiles or narrowband modulated spectrum echo samples for model training to ensure the stability and generalization ability of the recognition model.

[0003] However, in practical applications, especially for long-range, weak, or concealed targets, the acquired radar echo signals are inevitably severely contaminated by environmental noise, system thermal noise, and various clutter, resulting in a significant reduction in the signal-to-noise ratio (SNR). The presence of noise distorts the original time-frequency structure and statistical characteristics of the echo signal, causing traditional recognition models trained on high SNR data to experience a sharp decline in recognition performance, or even fail, when faced with low SNR measured data.

[0004] To address the challenge of low signal-to-noise ratio (SNR), existing technologies often employ a two-stage processing framework of "preprocessing followed by recognition." In the preprocessing stage, two main methods are used: First, classical signal denoising algorithms based on the spatial or transform domains. While these methods are intuitive in principle, they often have simple structures and fixed parameters, making it difficult to precisely preserve the target echo structure features crucial for recognition while suppressing noise, leading to the loss of effective information. Furthermore, their computational complexity may increase with signal dimensionality, hindering real-time processing. Second, deep learning-based denoising methods, such as using convolutional autoencoders or generative adversarial networks (GANs) for noise filtering. These methods outperform traditional methods in feature representation, but their model training objectives are usually limited to the fidelity of signal reconstruction, failing to achieve integrated, end-to-end joint optimization with subsequent recognition tasks. Therefore, while the visual effect of the signal preprocessed using these methods is improved, it may not be the optimal "clean" feature representation for target recognition, resulting in limited performance improvement.

[0005] In summary, existing technologies have significant shortcomings: traditional denoising methods struggle to balance noise suppression and feature preservation; while deep denoising methods are disconnected from the recognition model, failing to achieve globally optimal system performance. This has become a key technical bottleneck restricting the robust application of radar target recognition technology in harsh environments such as low signal-to-noise ratio and strong interference. Therefore, there is an urgent need for an integrated method that can deeply integrate prior signal knowledge with data-driven learning to achieve coordinated optimization of echo enhancement and target recognition. Summary of the Invention

[0006] To address the technical problems existing in the background art, this invention proposes an echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraints.

[0007] The echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraints proposed in this invention includes:

[0008] Acquire radar multimodal echo data of the target to be identified;

[0009] The radar multimode echo data is preprocessed to obtain the first radar multimode echo data;

[0010] The first radar multimodal echo data is input into a trained adaptive wavelet denoising and recognition integrated model to obtain the target recognition result corresponding to the radar multimodal echo data;

[0011] The adaptive wavelet denoising and recognition integrated model is used to perform signal structure decomposition and denoising on the input first radar multimodal echo data to enhance the echo, and to fuse and classify the enhanced echo features. The adaptive wavelet denoising and recognition integrated model includes an adaptive wavelet denoising module, a feature fusion module and a classification module connected in sequence.

[0012] Preferably, the adaptive wavelet denoising module is used to receive first radar multimodal echo data, perform signal structure decomposition and adaptive threshold denoising on the first radar multimodal echo data, and output the corresponding denoised broadband feature vector and denoised narrowband feature vector; the feature fusion module, whose input is connected to the output of the adaptive wavelet denoising module, is used to receive the denoised broadband feature vector and the denoised narrowband feature vector, and perform weighted fusion of the two types of feature vectors through an attention mechanism to generate a fused multimodal feature vector; the classification module, whose input is connected to the output of the feature fusion module, is used to receive the fused multimodal feature vector, perform fully connected layer mapping and classification decision on the fused multimodal feature vector, and output the final target category recognition result.

[0013] Preferably, the adaptive wavelet denoising module includes multiple cascaded time-frequency coding modules, skip connection layers corresponding to each time-frequency coding module, and decoding modules connected to each skip connection layer; the time-frequency coding module includes sequentially connected one-dimensional convolutional blocks and adaptive discrete wavelet transform layers; the decoding module includes sequentially connected upsampling layers, feature concatenation layers, and one-dimensional convolutional blocks; the skip connection layer includes at least one one-dimensional convolutional block; the one-dimensional convolutional block includes multiple sequentially connected one-dimensional convolutional layers and Leaky ReLU activation layers.

[0014] Preferably, the processing procedure of the adaptive discrete wavelet transform layer includes:

[0015] Obtain the low-pass and high-pass filters corresponding to the preset wavelet types;

[0016] The low-pass filter is used to filter the input features to obtain the low-frequency components;

[0017] The high-pass filter is used to filter the input features to obtain high-frequency components;

[0018] An adaptive threshold filtering algorithm is used to filter the high-frequency components to obtain the filtered high-frequency components;

[0019] The low-frequency component is added to the filtered high-frequency component to obtain the output features of the adaptive discrete wavelet transform layer.

[0020] Preferably, the training process of the adaptive wavelet denoising and recognition integrated model includes:

[0021] Obtain the original training dataset containing radar one-dimensional high-resolution range image samples and narrowband modulated spectrum echo samples;

[0022] Gaussian noise is added to the samples in the original training dataset to construct a noisy training dataset;

[0023] The samples in the training dataset are labeled to obtain their true class labels;

[0024] The initial adaptive wavelet denoising and recognition integrated model, which includes an adaptive wavelet denoising module, a feature fusion module, and a classification module, is iteratively trained using the training dataset. The model converges by jointly optimizing the reconstruction loss and the classification loss, thus obtaining the trained adaptive wavelet denoising and recognition integrated model.

[0025] Preferably, the joint optimization reconstruction loss and classification loss refer to the overall loss function of the adaptive wavelet denoising and recognition integrated model being the sum of broadband reconstruction loss, narrowband reconstruction loss and classification loss, and the training objective is to make the overall loss function converge.

[0026] Preferably, the formula for calculating the broadband reconstruction loss is as follows:

[0027] ;

[0028] in, Losses due to broadband reconstruction; The number of radar one-dimensional high-resolution range image samples in the training dataset; Indicates mean square error; For the training dataset, the first Broadband output features of a radar one-dimensional high-resolution range image sample after denoising; The first in the original training dataset A radar one-dimensional high-resolution range image sample; This indicates the search for the L2 norm.

[0029] Preferably, the formula for calculating the narrowband reconstruction loss is as follows:

[0030] ;

[0031] in, For narrowband reconstruction loss; The number of narrowband modulation spectrum echo samples in the training dataset; For the training dataset, the first Narrowband output characteristics after denoising of narrowband modulation spectrum echo samples; The first in the original training dataset A narrowband modulation spectrum echo sample.

[0032] Preferably, the formula for calculating the classification loss is as follows:

[0033] ;

[0034] in, For classification loss; The total number of data points in the training dataset; For the data index of the training dataset; For real labels, The predicted label is obtained by the adaptive wavelet denoising and recognition integrated model.

[0035] Preferably, the preprocessing of the radar multimodal echo data specifically includes:

[0036] Alignment and normalization are performed on the radar one-dimensional high-resolution range image data in the radar multi-mode echo data to obtain the preprocessed radar one-dimensional high-resolution range image data.

[0037] The narrowband modulation spectrum echo data in the radar multimode echo data is subjected to clutter removal, main body compensation and normalization to obtain preprocessed narrowband modulation spectrum echo data.

[0038] The first radar multimode echo data is composed of preprocessed one-dimensional high-resolution range image data and preprocessed narrowband modulation spectrum echo data.

[0039] This invention proposes an echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraints. By constructing an integrated adaptive wavelet denoising and identification model, it deeply integrates wavelet multi-resolution analysis from traditional signal processing with the feature learning capabilities of deep neural networks. This integrated adaptive wavelet denoising and identification model utilizes a learnable adaptive threshold for refined noise suppression in the wavelet domain, effectively filtering out noise while preserving the original echo structure features crucial for identification to the greatest extent. Through an end-to-end joint training strategy, the denoising process directly serves and optimizes the final identification task, overcoming the drawbacks of feature distortion and task disconnect in traditional staged processing. Simultaneously, the integrated adaptive wavelet denoising and identification model utilizes radar multimodal data for information complementarity and attention-based feature fusion, further enhancing the system's anti-interference capability and feature discrimination ability. This improves the accuracy, robustness, and practicality of radar target identification in complex and harsh environments such as low signal-to-noise ratio and strong interference. Attached Figure Description

[0040] Figure 1 This is a schematic diagram illustrating the workflow of the echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraints proposed in this invention.

[0041] Figure 2 This is a schematic diagram illustrating the implementation process of the echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraints proposed in this invention.

[0042] Figure 3 This is a schematic diagram of the implementation of the radar one-dimensional high-resolution range profile adaptive wavelet denoising module / narrowband modulation spectrum echo adaptive wavelet denoising module based on the signal structure constraint-based echo enhancement and low signal-to-noise ratio robust identification method proposed in this invention.

[0043] Figure 4 This is a schematic diagram of the architecture of the feature fusion module of the echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraints proposed in this invention.

[0044] Figure 5 This is a schematic diagram of the architecture of the classification module of the echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraints proposed in this invention. Detailed Implementation

[0045] Reference Figures 1-5 The echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraints proposed in this invention includes the following steps:

[0046] S1. Acquire radar multimode echo data of the target to be identified.

[0047] S2. Preprocess the radar multimode echo data to obtain the first radar multimode echo data.

[0048] In this embodiment, the radar multimode echo data is preprocessed, specifically including: aligning and normalizing the radar one-dimensional high-resolution range image data in the radar multimode echo data to obtain preprocessed radar one-dimensional high-resolution range image data; removing clutter, compensating for the main body, and normalizing the narrowband modulation spectrum echo data in the radar multimode echo data to obtain preprocessed narrowband modulation spectrum echo data; and using the preprocessed radar one-dimensional high-resolution range image data and the preprocessed narrowband modulation spectrum echo data to constitute the first radar multimode echo data.

[0049] It should be noted that the alignment and normalization of the one-dimensional high-resolution range image data of the radar to be identified in the radar multimodal echo data involves aligning and normalizing multiple data points within the one-dimensional high-resolution range image data of the radar to be identified. Clutter removal of the narrowband modulation spectrum echo data to be identified is to reduce the impact of clutter components on the identification process. Subject compensation is to mitigate the impact of spectral position changes caused by the movement of the target subject in the narrowband modulation spectrum echo data to be identified, thereby improving the usability of the data.

[0050] S3. Input the first radar multimodal echo data into the trained adaptive wavelet denoising and recognition integrated model to obtain the target recognition result corresponding to the radar multimodal echo data;

[0051] The adaptive wavelet denoising and recognition integrated model is used to perform signal structure decomposition and denoising on the input first radar multimodal echo data to enhance the echo, and to fuse and classify the enhanced echo features. The adaptive wavelet denoising and recognition integrated model includes an adaptive wavelet denoising module, a feature fusion module and a classification module connected in sequence.

[0052] In this embodiment, the adaptive wavelet denoising module receives first radar multimodal echo data, performs signal structure decomposition and adaptive threshold denoising on the first radar multimodal echo data, and outputs corresponding denoised broadband feature vectors and denoised narrowband feature vectors; the feature fusion module, whose input is connected to the output of the adaptive wavelet denoising module, receives the denoised broadband feature vectors and denoised narrowband feature vectors, and performs weighted fusion of the two types of feature vectors through an attention mechanism to generate a fused multimodal feature vector; the classification module, whose input is connected to the output of the feature fusion module, receives the fused multimodal feature vector, performs fully connected layer mapping and classification decision on the fused multimodal feature vector, and outputs the final target category recognition result.

[0053] Specifically, the adaptive wavelet denoising module includes multiple cascaded time-frequency coding modules, skip connection layers connected to each time-frequency coding module, and decoding modules connected to each skip connection layer. The time-frequency coding module includes sequentially connected one-dimensional convolutional blocks and adaptive discrete wavelet transform layers. The decoding module includes sequentially connected upsampling layers, feature concatenation layers, and one-dimensional convolutional blocks. Each skip connection layer includes at least one one-dimensional convolutional block. Each one-dimensional convolutional block includes multiple sequentially connected one-dimensional convolutional layers and Leaky ReLU activation layers.

[0054] In this embodiment, the skip connection layer includes at least one 1×1 one-dimensional convolutional block; the one-dimensional convolutional block includes multiple sequentially connected 1×3 one-dimensional convolutional layers and Leaky ReLU activation layers.

[0055] In this embodiment, the processing procedure of the adaptive discrete wavelet transform layer includes: obtaining a low-pass filter and a high-pass filter corresponding to a preset wavelet type; filtering the input features using the low-pass filter to obtain low-frequency components; filtering the input features using the high-pass filter to obtain high-frequency components; filtering the high-frequency components using an adaptive threshold filtering algorithm to obtain filtered high-frequency components; and adding the low-frequency components and the filtered high-frequency components to obtain the output features of the adaptive discrete wavelet transform layer.

[0056] In the implementation, the filter parameters are related to the wavelet type, and the processing procedure of the adaptive discrete wavelet transform layer can be referred to as follows:

[0057] (1) Obtain the low-pass filter corresponding to the adaptive discrete wavelet transform layer and high-pass filter .

[0058] (2) The input features of the adaptive discrete wavelet transform layer are respectively passed through a low-pass filter. and high-pass filter ,in, This represents the value of the low-pass filter; This represents the value of the high-pass filter; it decomposes the input features into low-frequency components. and high-frequency components containing noise and a small amount of cleaning content. , means as follows:

[0059]

[0060] in, Indicates the first Input features of the layer adaptive discrete wavelet transform layer Show convolutional layers.

[0061] (3) Using adaptive threshold filtering for high-frequency components Perform filtering and output the filtered high-frequency components. ;

[0062] It is worth mentioning that the thresholding denoising strategy includes various threshold functions such as soft thresholding and hard thresholding; among them, the threshold parameter... It is obtained through parameterized learning of the integrated adaptive wavelet denoising and recognition model, and the formula is expressed as follows:

[0063] ;

[0064] in, Indicates the ReLU activation layer; Indicates a linear layer; The initial value can be preset to 0.1.

[0065] (4) Utilizing low-frequency components With the filtered high-frequency components Summing yields the output of the adaptive discrete wavelet transform layer:

[0066] ;

[0067] In scenarios with extremely low signal-to-noise ratios, high-frequency components can be removed to improve the output of the adaptive discrete wavelet transform layer. .

[0068] Specifically, the training process of the adaptive wavelet denoising and recognition integrated model includes: acquiring an original training dataset containing radar one-dimensional high-resolution range image samples and narrowband modulation spectrum echo samples; adding Gaussian noise to the samples in the original training dataset to construct a noisy training dataset; labeling the samples in the training dataset to obtain their true class labels; using the training dataset to iteratively train the initial adaptive wavelet denoising and recognition integrated model containing an adaptive wavelet denoising module, a feature fusion module, and a classification module; and converging the model by jointly optimizing the reconstruction loss and the classification loss to obtain the trained adaptive wavelet denoising and recognition integrated model.

[0069] It should be noted that the joint optimization reconstruction loss and classification loss refer to the fact that the overall loss function of the adaptive wavelet denoising and recognition integrated model is the sum of the broadband reconstruction loss, the narrowband reconstruction loss and the classification loss, and the training objective is to make this overall loss function converge.

[0070] Specifically, Gaussian noise is added independently to each radar one-dimensional high-resolution range image sample and narrowband modulation spectrum echo sample in the original training dataset to obtain the training dataset. The signal-to-noise ratio of each data in the training dataset can range from -10dB to 30dB.

[0071] In this embodiment, the adaptive wavelet denoising module includes a radar one-dimensional high-resolution range image adaptive wavelet denoising module and a narrowband modulation spectrum echo adaptive wavelet denoising module.

[0072] In this embodiment, the preprocessed one-dimensional high-resolution range profile data of the radar to be identified from the preprocessed radar multimodal echo data is input into the radar one-dimensional high-resolution range profile adaptive wavelet denoising module to obtain the denoised broadband output features; based on the broadband output features and the preprocessed one-dimensional high-resolution range profile data of the radar to be identified, the calculation formula for the broadband reconstruction loss is as follows:

[0073] ;

[0074] in, Losses due to broadband reconstruction; The number of radar one-dimensional high-resolution range image samples in the training dataset; Indicates mean square error; For the training dataset, the first Broadband output features of a radar one-dimensional high-resolution range image sample after denoising; The first in the original training dataset A radar one-dimensional high-resolution range image sample; This indicates the search for the L2 norm.

[0075] In this embodiment, the preprocessed narrowband modulation spectrum echo data to be identified from the preprocessed radar multimodal echo data is input into the narrowband modulation spectrum echo adaptive wavelet denoising module to obtain the denoised narrowband output features; based on the narrowband output features and the preprocessed narrowband modulation spectrum echo data to be identified, the specific formula for calculating the narrowband reconstruction loss is as follows:

[0076] ;

[0077] in, For narrowband reconstruction loss; The number of narrowband modulation spectrum echo samples in the training dataset; For the training dataset, the first Narrowband output characteristics after denoising of narrowband modulation spectrum echo samples; The first in the original training dataset A narrowband modulation spectrum echo sample.

[0078] In this embodiment, the denoised broadband output features and the denoised narrowband output features are simultaneously input into the feature fusion module for processing, and then input into the classification module for recognition to obtain the recognition result. Based on the classification result and the true labels of the training data in the training dataset, the classification loss is calculated. The specific formula for calculating the classification loss is as follows:

[0079] ;

[0080] in, For classification loss; The total number of data points in the training dataset; For the data index of the training dataset; For real labels, The predicted label is obtained by the adaptive wavelet denoising and recognition integrated model.

[0081] like Figure 2 As shown, when processing radar multimodal echo data using the integrated adaptive wavelet denoising and recognition model, the one-dimensional high-resolution range profile data and the narrowband modulation spectrum echo data of the radar to be identified are preprocessed separately. After the preprocessed radar multimodal echo data is input into the integrated adaptive wavelet denoising and recognition model, the one-dimensional high-resolution range profile data and the narrowband modulation spectrum echo data of the radar to be identified are respectively input into the adaptive wavelet denoising module for the radar one-dimensional high-resolution range profile and the narrowband modulation spectrum echo, respectively, to obtain the output features of the one-dimensional high-resolution range profile data and the output features of the narrowband modulation spectrum echo. The obtained output features of the one-dimensional high-resolution range profile data and the output features of the narrowband modulation spectrum echo are then input into the feature fusion module for multimodal feature fusion and target classification based on the attention mechanism. Finally, the recognition result corresponding to the radar multimodal echo data is obtained according to the classification module.

[0082] like Figure 3 As shown, the adaptive wavelet denoising module includes multiple time-frequency coding modules connected in sequence. Each time-frequency coding module is connected to a corresponding skip connection layer, and each skip connection layer is connected to a decoding module. The time-frequency coding module includes a one-dimensional convolutional block (1-D Conv) and an adaptive discrete wavelet transform (DWT) layer connected in sequence. The decoding module includes an upsampling layer, a feature concatenation layer, and a one-dimensional convolutional block (1-D Conv) connected in sequence. The skip connection layer includes multiple one-dimensional convolutional blocks (1-D Conv) connected in sequence. The one-dimensional convolutional block (1-D Conv) includes multiple one-dimensional convolutional layers and a Leaky ReLU layer connected in sequence.

[0083] In addition to the adaptive wavelet denoising module, the feature fusion module includes a feature concatenation layer and an attention layer connected in sequence. A preferred embodiment architecture is as follows: Figure 4As shown, the denoised radar one-dimensional high-resolution range image and narrowband modulation spectrum echo are respectively passed through three sequentially connected one-dimensional convolutional layers (1×3 Conv1d) and Leaky ReLU layers to obtain radar one-dimensional high-resolution range image features and narrowband modulation spectrum echo features. After feature concatenation, attention weights are obtained through an average pooling layer (Avgpool), a linear layer (Linear), a Leaky ReLU layer, a linear layer (Linear), and a Sigmoid layer. The attention weights are multiplied by the concatenated features to obtain the fused features. The classification module includes multiple linear layers, and the preferred embodiment architecture is as follows. Figure 5 As shown, the fused features are sequentially processed through a flattened layer, a linear layer, a Leaky ReLU layer, and another linear layer to obtain the recognition result.

[0084] In the implementation method, the effectiveness of the target fusion and recognition method based on radar multimodal data provided in the above embodiments is verified through simulation experiments:

[0085] (1) Simulation conditions:

[0086] Simulation test conditions: The processor is an Intel(R) i7 CPU with a clock speed of 3.20GHz, the memory capacity is 64GB, the graphics card is an Nvidia RTX 3090 with a video memory capacity of 24GB.

[0087] The software platform for the simulation experiment in this embodiment is Python 3.8 and PyTorch 2.0.

[0088] (2) Simulation content and result analysis:

[0089] In the simulation experiments of this application, the method provided in the above embodiments and a prior art are used to conduct target denoising and recognition experiments on three types of aircraft, and the recognition rate under low signal-to-noise ratio is compared. Gaussian white noise of 0dB to 30dB is added to the complex domain echo of the original training dataset as the training dataset, and the recognition rate of the test set under a 10dB scene is tested. The prior art refers to the paper "U-Net: Convolutional Networks for Biomedical Image Segmentation". The original network is designed for two-dimensional image data. In this simulation experiment, its main structure is retained, and the two-dimensional convolutional layers are changed to one-dimensional convolutional layers to adapt to one-dimensional signal input. The number of channels is consistent with the number in the embodiments of this application. The recognition rate is introduced to evaluate the performance of the recognition model. The recognition rate is defined as the ratio of the number of correctly recognized samples to the total number of test samples. The higher the recognition rate, the more correctly recognized samples there are, and the better the recognition performance of the model. The recognition rate comparison results are shown in Table 1.

[0090] Table 1: Recognition Rate Comparison Table

[0091]

[0092] By comparing the recognition rates of this application and existing technologies, it can be seen that the signal structure constraints used in this application can effectively achieve echo enhancement. Combined with the multi-mode data fusion strategy, the final recognition performance is greatly improved compared with existing methods, thus enhancing the radar target recognition performance in low signal-to-noise ratio scenarios.

[0093] In the above technical solution, this application constructs an integrated adaptive wavelet denoising and recognition model, which combines traditional signal processing technology and deep network noise suppression technology to enhance the model's ability to enhance and robustly recognize low signal-to-noise ratio radar echoes. Furthermore, this application utilizes radar multimodal echo data for recognition, supplements information through data from different modes, comprehensively extracts and utilizes target characteristics, enhances the robustness of the model, and improves the target recognition accuracy and stability.

[0094] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for echo enhancement and robust identification of low signal-to-noise ratio based on signal structure constraints, characterized in that, include: Acquire radar multimodal echo data of the target to be identified; The radar multimode echo data is preprocessed to obtain the first radar multimode echo data; The first radar multimodal echo data is input into a trained adaptive wavelet denoising and recognition integrated model to obtain the target recognition result corresponding to the radar multimodal echo data; The adaptive wavelet denoising and recognition integrated model is used to perform signal structure decomposition and denoising on the input first radar multimodal echo data to enhance the echo, and to fuse and classify the enhanced echo features. The adaptive wavelet denoising and recognition integrated model includes an adaptive wavelet denoising module, a feature fusion module and a classification module connected in sequence. The adaptive wavelet denoising module receives multimodal echo data from the first radar, performs signal structure decomposition and adaptive threshold denoising on the multimodal echo data, and outputs corresponding denoised wideband feature vectors and denoised narrowband feature vectors. The feature fusion module, whose input is connected to the output of the adaptive wavelet denoising module, receives the denoised wideband feature vectors and denoised narrowband feature vectors, and performs weighted fusion of the two types of feature vectors through an attention mechanism to generate a fused multimodal feature vector. The classification module, whose input is connected to the output of the feature fusion module, receives the fused multimodal feature vector, performs fully connected layer mapping and classification decision on the fused multimodal feature vector, and outputs the final target category recognition result. The adaptive wavelet denoising module includes multiple cascaded time-frequency coding modules, skip connection layers corresponding to each time-frequency coding module, and decoding modules connected to each skip connection layer. Each time-frequency coding module includes sequentially connected one-dimensional convolutional blocks and adaptive discrete wavelet transform layers. The decoding module includes sequentially connected upsampling layers, feature concatenation layers, and one-dimensional convolutional blocks. Each skip connection layer includes at least one one-dimensional convolutional block. Each one-dimensional convolutional block includes multiple sequentially connected one-dimensional convolutional layers and Leaky ReLU activation layers. The processing procedure of the adaptive discrete wavelet transform layer includes: Obtain the low-pass and high-pass filters corresponding to the preset wavelet types; The low-pass filter is used to filter the input features to obtain the low-frequency components; The high-pass filter is used to filter the input features to obtain high-frequency components; An adaptive threshold filtering algorithm is used to filter the high-frequency components to obtain the filtered high-frequency components; The low-frequency component is added to the filtered high-frequency component to obtain the output features of the adaptive discrete wavelet transform layer; The training process of the adaptive wavelet denoising and recognition integrated model includes: Obtain the original training dataset containing radar one-dimensional high-resolution range image samples and narrowband modulated spectrum echo samples; Gaussian noise is added to the samples in the original training dataset to construct a noisy training dataset; The samples in the training dataset are labeled to obtain their true class labels; The initial adaptive wavelet denoising and recognition integrated model, which includes an adaptive wavelet denoising module, a feature fusion module, and a classification module, is iteratively trained using the training dataset. The model converges by jointly optimizing the reconstruction loss and the classification loss, thus obtaining the trained adaptive wavelet denoising and recognition integrated model.

2. The echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraints according to claim 1, characterized in that, The joint optimization reconstruction loss and classification loss refer to the fact that the overall loss function of the adaptive wavelet denoising and recognition integrated model is the sum of the broadband reconstruction loss, the narrowband reconstruction loss and the classification loss, and the training objective is to make the overall loss function converge.

3. The echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraints according to claim 2, characterized in that, The specific formula for calculating the broadband reconstruction loss is as follows: ； in, This is due to the loss from broadband reconstruction. The number of radar one-dimensional high-resolution range image samples in the training dataset; Indicates mean square error; For the training dataset, the first Broadband output features of a radar one-dimensional high-resolution range image sample after denoising; The first in the original training dataset A radar one-dimensional high-resolution range image sample; This indicates the search for the L2 norm.

4. The echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraints according to claim 2, characterized in that, The specific formula for calculating the narrowband reconstruction loss is as follows: ； in, For narrowband reconstruction loss; The number of narrowband modulation spectrum echo samples in the training dataset; For the training dataset, the first Narrowband output characteristics after denoising of narrowband modulation spectrum echo samples; The first in the original training dataset A narrowband modulation spectrum echo sample.

5. The echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraints according to claim 2, characterized in that, The specific formula for calculating the classification loss is as follows: ； in, For classification loss; The total number of data points in the training dataset; For the data index of the training dataset; For real labels, The predicted label is obtained by the adaptive wavelet denoising and recognition integrated model.

6. The echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraints according to claim 1, characterized in that, The preprocessing of radar multimodal echo data specifically includes: Alignment and normalization are performed on the radar one-dimensional high-resolution range image data in the radar multi-mode echo data to obtain the preprocessed radar one-dimensional high-resolution range image data. The narrowband modulation spectrum echo data in the radar multimode echo data is subjected to clutter removal, main body compensation and normalization to obtain preprocessed narrowband modulation spectrum echo data. The first radar multimode echo data is composed of preprocessed one-dimensional high-resolution range image data and preprocessed narrowband modulation spectrum echo data.

Images