Radar weak target detection method and device based on deep learning

Abstract

The invention relates to the technical field of radar weak target detection, in particular to a radar weak target detection method and device based on deep learning. According to the technical scheme, the method comprises the following steps: data preprocessing and enhancement: receiving an original echo signal of a radar, generating three-dimensional time frequency-wavelet coefficient spatial data by combining time frequency analysis with wavelet transform, and expanding a data set in multiple modes; and feature extraction network construction: constructing a 3D-CNN and LSTM fused network, wherein the 3D-CNN has a convolution operation with a variable convolution kernel size. According to the method, the data are richer and more comprehensive through time-frequency analysis in combination with wavelet transform and data enhancement, and multi-dimensional features can be accurately captured through 3D-CNN and LSTM fusion and an adaptive convolution kernel; and the training effect is ensured through the adaptive learning rate and the early stop method, and the accuracy and adaptability of weak target detection are jointly improved.

Description

Technical Field

[0001] The present invention relates to the field of radar weak target detection technology, and in particular to a radar weak target detection method and a device based on deep learning. Background Art

[0002] In radar weak target detection, a significant problem is that in a low signal-to-noise ratio environment, the echo signal of a weak target is easily submerged by noise, making it difficult to effectively extract target features. Traditional signal processing methods, such as matched filtering and constant false alarm rate detection, can work well in high signal-to-noise ratio scenarios, but their performance will drop significantly at low signal-to-noise ratios. This is because the echo signal energy of a weak target is low, and the distinction between it and noise in time domain, frequency domain and other features is not high, making it difficult for traditional methods to accurately distinguish target signals and noise, and thus unable to accurately extract target characteristic information such as distance, speed, and amplitude. Therefore, the present application proposes a radar weak target detection method and device based on deep learning. Summary of the invention

[0003] The purpose of the present invention is to propose a radar weak target detection method and device based on deep learning to address the problem of difficulty in extracting target features under low signal-to-noise ratio in the background technology.

[0004] On the one hand, the present application provides a radar weak target detection method based on deep learning, comprising the following method steps:

[0005] Data preprocessing and enhancement: Receive the original radar echo signal, use time-frequency analysis combined with wavelet transform to generate three-dimensional time-frequency-wavelet coefficient space data, and expand the data set in various ways;

[0006] Feature extraction network construction: Build a network that integrates 3D-CNN and LSTM. 3D-CNN has convolution operations with variable convolution kernel sizes.

[0007] Feature fusion and target detection: Use a module based on a multimodal attention mechanism to fuse features, and input the fused features into a target detection module based on GAN;

[0008] Model training and optimization: Train the network with labeled datasets, using adaptive learning rate adjustment algorithm and early stopping method;

[0009] Real-time detection and result output: The real-time echo signal is pre-processed and input into the trained model, and the target information is output and visualized, with special display for weak targets.

[0010] Optionally, the data processing and enhancement specifically includes: receiving the original radar echo signal, processing the signal by using time-frequency analysis combined with wavelet transform, and generating three-dimensional time-frequency-wavelet coefficient space data, wherein the time-frequency analysis is used to convert the time domain signal into a time-frequency domain representation, and the wavelet transform is used to further highlight the characteristics of the signal at different scales;

[0011] The generated 3D data is enhanced by rotating, flipping, adding random noise, and simulating different target motion trajectories to expand the data set so that the data set contains more types of weak target signals and noise mixtures.

[0012] The specific formula of the time-frequency analysis combined with wavelet transform is: Assuming the radar echo signal is s(τ), its short-time Fourier transform STFT(t,f) is expressed as:

[0013]

[0014] Where ω(τ) is the window function;

[0015] Wavelet transform WT(a,b) is expressed as:

[0016]

[0017] Among them, a is the scale factor, b is the translation factor, and ψ(t) is the wavelet basis function; the combined three-dimensional data is expressed as D(t,f,a)=STFT(t,f)×WT(a,b), where b is dynamically determined according to the signal characteristics and the target motion state.

[0018] Optionally, the feature extraction network construction specifically includes the following steps:

[0019] Construct a feature extraction network that integrates a three-dimensional convolutional neural network and a long short-term memory network, where 3D-CNN is used to extract local features of three-dimensional data in the spatial dimension, and LSTM is used to capture the dynamic change characteristics of the signal in the temporal dimension;

[0020] In the convolution layer of 3D-CNN, a convolution operation with a variable convolution kernel size is designed to adaptively adjust the convolution kernel size according to the complexity of the signal in different time-frequency-wavelet coefficient regions;

[0021] The convolution operation with variable convolution kernel size is performed at the convolution kernel size K of the i-th convolution layer. i The calculation formula is:

[0022]

[0023] Among them, K0 is the basic convolution kernel size, α is the adjustment coefficient, M is the number of channels of the input feature map, and c i-1(j) is the feature map of the jth channel in the previous layer.

[0024] Optionally, the feature fusion and target detection specifically include:

[0025] After the feature extraction network, a feature fusion module based on the multimodal attention mechanism is designed. This module can dynamically assign weights according to the importance of different feature modalities to target detection and fuse the features extracted by 3D-CNN and LSTM.

[0026] The fused features are input into a target detection module based on a generative adversarial network (GAN). The generator is used to generate possible target signal hypotheses, and the discriminator is used to judge the similarity between the generated hypotheses and the real target signals. The accuracy of target detection is improved through adversarial training.

[0027] In the feature fusion module based on the multimodal attention mechanism, the attention weight w mn The calculation formula is:

[0028]

[0029] Among them, e mn It is the associated energy value of the mth 3D-CNN feature and the nth LSTM feature calculated by a bilinear mapping function. P and Q are the number of 3D-CNN and LSTM output features respectively. The calculation formula of the fused feature F is:

[0030]

[0031] Among them, F 3D-CNN (m) is the mth feature output by 3D-CNN, F LSTM (n) is the nth feature output by LSTM.

[0032] Optionally, the model training and optimization specifically includes the following steps:

[0033] The entire network is trained using a large number of annotated data sets containing weak target signals. During the training process, an adaptive learning rate adjustment algorithm is used to dynamically adjust the learning rate according to the loss changes of the model on the validation set.

[0034] Use the early stopping method to prevent overfitting. When the performance indicators of the model on the validation set do not improve for several consecutive training cycles, stop training and save the current optimal model.

[0035] The adaptive learning rate adjustment algorithm is: assuming that the current learning rate is η, the training cycle is t, and the validation set loss is L(t), then the learning rate η(t+1) of the next training cycle is calculated as follows:

[0036]

[0037] Among them, β is the learning rate adjustment factor, when η(t+1)>η max When η(t+1)=η max , when η(t+1)<η min When η(t+1)=η min , η max and η min are the upper and lower limits of the learning rate.

[0038] Optionally, the number of consecutive training cycles in the early stopping method is N, and N is dynamically determined according to the model complexity and the size of the data set, and the value range is between 5-15.

[0039] Optionally, the real-time detection and result output are specifically:

[0040] The radar echo signal received in real time is input into the trained model after data preprocessing. The model outputs the judgment result of whether the target exists and the characteristic information of the target, including distance, speed, amplitude and target movement trend;

[0041] The output results are visualized and the target information is intuitively displayed on the radar screen through different colors and marks. A special highlighting method is used for weak targets to facilitate identification.

[0042] Optionally, the special highlighting mode is to use a flickering effect and increase the target mark size, and the flickering frequency and the mark size increase ratio are adaptively adjusted according to the faintness of the target, and the faintness is measured by the target signal-to-noise ratio.

[0043] Optionally, the annotation information of the annotated data set containing weak target signals includes the actual existence of the target, the precise distance, speed, amplitude of the target, and the starting point and end point of the target motion trajectory, and the annotation is completed by combining high-precision radar measurement equipment with manual assisted annotation.

[0044] On the other hand, the present application provides a radar weak target detection device based on deep learning, which adopts the radar weak target detection method based on deep learning described in the first aspect above, including:

[0045] Signal acquisition module: used to receive radar echo signals and transmit them to the data processing module;

[0046] Data processing module: connected to the signal acquisition module, including a time-frequency analysis combined with a wavelet transform unit and a data enhancement unit. The time-frequency analysis combined with a wavelet transform unit processes the received radar echo signal into three-dimensional time-frequency-wavelet coefficient space data. The data enhancement unit rotates, flips, adds random noise to the three-dimensional data, and simulates the motion trajectories of different targets.

[0047] Feature extraction module: connected to the data processing module, including a three-dimensional convolutional neural network unit and a long short-term memory network unit. The 3D-CNN unit has the function of adaptively adjusting the size of the convolution kernel according to the complexity of the signal in different time-frequency-wavelet coefficient regions, and is used to extract the local features of three-dimensional data in the spatial dimension. The LSTM unit is used to capture the dynamic change characteristics of the signal in the time dimension;

[0048] Feature fusion and target detection module: connected to the feature extraction module, including a feature fusion unit based on a multimodal attention mechanism and a target detection unit based on a generative adversarial network. The feature fusion unit dynamically allocates weights according to the importance of different feature modalities to target detection to fuse the features extracted by 3D-CNN and LSTM. The target detection unit detects targets by generating possible target signal hypotheses through a generator and judging the similarity between the hypotheses and the real target signals through a discriminator.

[0049] Model training and optimization module: connected to the feature fusion and target detection module, including an adaptive learning rate adjustment unit and an early stopping control unit. The adaptive learning rate adjustment unit dynamically adjusts the learning rate according to the loss change of the model on the validation set, and the early stopping control unit stops training when the performance index of the model on the validation set does not improve for a specific training cycle;

[0050] Display module: connected to the model training and optimization module, used to receive the judgment result of the model output on the existence of the target and the target's distance, speed, amplitude and target movement trend characteristic information, and visualize the target information. For faint targets, a special highlight display method is used. The highlight display method includes a flashing effect and an increase in the target mark size. The flashing frequency and the mark size increase ratio are adaptively adjusted according to the faintness of the target. The faintness is measured by the target signal-to-noise ratio.

[0051] Compared with the prior art, the present invention has at least one of the following beneficial technical effects:

[0052] 1. The present invention combines time-frequency analysis with wavelet transform to generate three-dimensional time-frequency-wavelet coefficient space data, which provides a multi-dimensional perspective for subsequent processing. The data representation method can highlight it, avoiding the loss of information under the traditional single data representation method, thus laying the foundation for more accurate target detection. Through a variety of data enhancement methods, various changes in target signals are covered, so that the model can be exposed to a large number of weak target signals in different situations during the training process, which greatly enhances the adaptability of the model to various actual scenarios and reduces the risk of overfitting due to data limitations.

[0053] 2. The network structure of 3D-CNN and LSTM fusion realizes the coordinated extraction of spatial and temporal dimension features. 3D-CNN extracts local features of three-dimensional data in the spatial dimension, and can capture the local texture and structural information of weak targets in the time-frequency-wavelet coefficient space. Time-related features such as phase change and frequency drift of target signals can be effectively extracted. The combination of the two provides strong support for a comprehensive understanding of weak target features. For areas with complex signals, larger convolution kernels can obtain more extensive contextual information, which helps to understand the complex target feature structure; while in areas with relatively simple signals, smaller convolution kernels can extract local features more finely. It overcomes the shortcomings of traditional fixed convolution kernels in processing complex and changeable signals, and improves the accuracy and pertinence of feature extraction.

[0054] 3. The feature fusion module based on the multimodal attention mechanism dynamically assigns weights according to the importance of the features to target detection, avoiding the problem of over-amplifying noise features or irrelevant features that may occur in traditional fusion methods, so that the fused features can better represent the essence of the target, thereby improving the accuracy of target detection, especially when weak target signals are easily interfered by noise, which can highlight its advantages.

[0055] 4. The GAN-based target detection module improves target detection performance through adversarial training of the generator and the discriminator. The target signal hypothesis generated by the generator can help the model explore more possible target feature patterns, and the discriminator continuously optimizes its judgment ability by comparing with the real target signal, more accurately distinguishing between targets and noise, and improving the sensitivity and specificity of target detection, which is difficult to achieve with traditional target detection methods.

[0056] 5. The adaptive learning rate adjustment algorithm dynamically adjusts the learning rate according to the change in the model's loss on the validation set. The adaptive mechanism can make model training more stable and efficient, and make full use of training data, especially when dealing with complex weak target detection problems, which helps to improve the final performance of the model. When the performance indicators of the model on the validation set have not improved for several consecutive training cycles, the training is stopped, which effectively prevents the model from overfitting the training data. In the weak target detection scenario, data is often hard to come by and may have certain limitations. Overfitting will seriously affect the generalization ability of the model in practical applications. The application of the early stopping method ensures that the model stops at the appropriate training stage, so that the model can maintain good generalization performance under limited data conditions, and improves the applicability of the model in different radar systems and actual environments.

[0057] In summary, the present invention combines time-frequency analysis with wavelet transform and data enhancement to make the data richer and more comprehensive, and can accurately capture multi-dimensional features through 3D-CNN and LSTM fusion and adaptive convolution kernel; improve the sensitivity and specificity of target detection, ensure the training effect through adaptive learning rate and early stopping method, and jointly improve the accuracy and adaptability of weak target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of the radar weak target detection method based on deep learning;

[0059] Figure 2 This is the principle block diagram of the radar weak target detection device based on deep learning. DETAILED DESCRIPTION

[0060] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0061] Example 1

[0062] like Figure 1 The radar weak target detection method based on deep learning includes five parts: data preprocessing and enhancement, feature extraction network construction, feature fusion and target detection, model training and optimization, real-time detection and result output. Each part is described in detail below.

[0063] Data preprocessing and enhancement: Receive the original radar echo signal, process the signal using time-frequency analysis combined with wavelet transform, and generate three-dimensional time-frequency-wavelet coefficient space data. Time-frequency analysis is used to convert the time domain signal into time-frequency domain representation, and wavelet transform is used to further highlight the characteristics of the signal at different scales.

[0064] The generated 3D data is enhanced by rotating, flipping, adding random noise, and simulating different target motion trajectories to expand the data set so that the data set contains more types of weak target signals and noise mixtures.

[0065] The specific formula of the time-frequency analysis combined with wavelet transform is: Assuming the radar echo signal is s(τ), its short-time Fourier transform STFT(t,f) is expressed as:

[0066]

[0067] Where ω(τ) is the window function;

[0068] Wavelet transform WT(a,b) is expressed as:

[0069]

[0070] Among them, a is the scale factor, b is the translation factor, and ψ(t) is the wavelet basis function; the combined three-dimensional data is expressed as D(t,f,a)=STFT(t,f)×WT(a,b), where b is dynamically determined according to the signal characteristics and the target motion state.

[0071] The radar echo signal is converted into three-dimensional time-frequency-wavelet coefficient space data by combining time-frequency analysis with wavelet transform. This data form can more comprehensively preserve the time-frequency and multi-scale characteristics of the signal. Compared with the traditional method of using only time-frequency analysis or a single wavelet transform, it provides richer information for subsequent feature extraction and target detection, and can mine the hidden features of weak targets in complex noise environments, which is impossible to achieve with traditional methods. Feature extraction network construction: Feature extraction network construction specifically includes the following steps:

[0072] Construct a feature extraction network that integrates a three-dimensional convolutional neural network and a long short-term memory network, where 3D-CNN is used to extract local features of three-dimensional data in the spatial dimension, and LSTM is used to capture the dynamic change characteristics of the signal in the temporal dimension;

[0073] In the convolution layer of 3D-CNN, a convolution operation with a variable convolution kernel size is designed to adaptively adjust the convolution kernel size according to the complexity of the signal in different time-frequency-wavelet coefficient regions;

[0074] Convolution operations with variable kernel sizes have kernel size K in the i-th convolutional layer.i The calculation formula is:

[0075]

[0076] Among them, K0 is the basic convolution kernel size, α is the adjustment coefficient, M is the number of channels of the input feature map, and c i-1 (j) is the feature map of the jth channel in the previous layer.

[0077] The fusion of 3D-CNN and LSTM and the design of variable convolution kernel size in 3D-CNN are extremely creative. The combination of 3D-CNN and LSTM can capture the characteristics of the signal in both spatial and temporal dimensions, and the variable convolution kernel size can be adaptively adjusted according to the signal characteristics, solving the problem of poor adaptability of traditional fixed convolution kernels when processing complex and changeable weak target signals. This makes feature extraction more accurate and can effectively meet the detection needs of different types of weak targets.

[0078] Feature fusion and target detection, including:

[0079] After the feature extraction network, a feature fusion module based on the multimodal attention mechanism is designed. This module can dynamically assign weights according to the importance of different feature modalities to target detection and fuse the features extracted by 3D-CNN and LSTM.

[0080] The fused features are input into a target detection module based on a generative adversarial network (GAN). The generator is used to generate possible target signal hypotheses, and the discriminator is used to judge the similarity between the generated hypotheses and the real target signals. The accuracy of target detection is improved through adversarial training.

[0081] In the feature fusion module based on the multimodal attention mechanism, the attention weight w mn The calculation formula is:

[0082]

[0083] Among them, e mn It is the associated energy value of the mth 3D-CNN feature and the nth LSTM feature calculated by a bilinear mapping function. P and Q are the number of 3D-CNN and LSTM output features respectively. The calculation formula of the fused feature F is:

[0084]

[0085] Among them, F 3D-CNN (m) is the mth feature output by 3D-CNN, F LSTM(n) is the nth feature of the LSTM output. The combination of the feature fusion module based on the multimodal attention mechanism and the target detection module based on the GAN is a novel design. The attention mechanism dynamically assigns weights according to the importance of the features to target detection, allowing the model to focus on key features and avoid the limitation of treating all features equally. The application of GAN further improves the accuracy of target detection through adversarial training of the generator and the discriminator, especially when weak target signals are similar to noise, and can better distinguish between the two, which is difficult to break through with existing technologies.

[0086] Model training and optimization include the following steps:

[0087] The entire network is trained using a large number of annotated data sets containing weak target signals. During the training process, an adaptive learning rate adjustment algorithm is used to dynamically adjust the learning rate according to the loss changes of the model on the validation set.

[0088] Use the early stopping method to prevent overfitting. When the performance indicators of the model on the validation set do not improve for several consecutive training cycles, stop training and save the current optimal model.

[0089] The adaptive learning rate adjustment algorithm is: let the current learning rate be η, the training cycle be t, and the validation set loss be L(t), then the learning rate η(t+1) of the next training cycle is calculated as follows:

[0090]

[0091] Among them, β is the learning rate adjustment factor, when η(t+1)>η max When η(t+1)=η max , when η(t+1)<η min When η(t+1)=η min , η max and η min are the upper and lower limits of the learning rate.

[0092] The number of consecutive training cycles in the early stopping method is N, which is dynamically determined according to the model complexity and the size of the data set, and the value range is between 5-15. The network is trained with a labeled data set, and an adaptive learning rate adjustment algorithm and early stopping method are used;

[0093] Real-time detection and result output: The real-time echo signal is pre-processed and input into the trained model, and the target information is output and visualized, with special display for faint targets. The combination of the adaptive learning rate adjustment algorithm and the early stopping method plays an important role in the model training process. The adaptive learning rate is dynamically adjusted according to the loss of the model on the validation set, making the model training more stable and efficient. The early stopping method can prevent overfitting and ensure that the model stops at the appropriate training stage, thereby improving the generalization ability of the model. This training and optimization method can make full use of limited data sets and improve the performance of the model in different environments.

[0094] It is worth noting that the real-time detection and result output are as follows:

[0095] The radar echo signal received in real time is input into the trained model after data preprocessing. The model outputs the judgment result of whether the target exists and the characteristic information of the target, including distance, speed, amplitude and target movement trend;

[0096] The output results are visualized and the target information is intuitively displayed on the radar screen through different colors and marks. A special highlighting method is used for weak targets to facilitate identification.

[0097] The special highlighting method is to use a flashing effect and increase the target mark size. The flashing frequency and the mark size increase ratio are adaptively adjusted according to the faintness of the target. The faintness is measured by the target signal to noise ratio.

[0098] The labeled data set containing weak target signals includes the actual existence of the target, the target's precise distance, speed, amplitude, and the starting and ending points of the target's motion trajectory. The labeling is completed by combining high-precision radar measurement equipment with manual assisted labeling.

[0099] The present invention can still maintain a high detection performance in an extremely low signal-to-noise ratio environment. Weak target signals that were completely submerged by noise in traditional methods can be accurately detected through this innovative method and device, greatly expanding the limit conditions for radar weak target detection. And it can effectively handle weak targets with complex motion trajectories. For example, for weak targets that suddenly change direction, change speed, or have irregular motion patterns, the system can accurately track and detect them. This is because multiple target motion trajectories are simulated during the data enhancement process, and the model can capture the dynamic change characteristics of the signal in the time dimension. This detection capability for complex moving targets exceeds the expectations of general technology.

[0100] Furthermore, when faced with new radar system parameters, different types of weak targets, or new noise environments, the model can quickly adapt and maintain good detection performance. This is due to the adaptive learning rate adjustment and data enhancement technology in model training, which enables the model to achieve effective weak target detection in new scenarios without a lot of retraining.

[0101] Example 2

[0102] like Figure 2 As shown, this embodiment provides a radar weak target detection device based on deep learning, including:

[0103] Signal acquisition module: used to receive radar echo signals and transmit them to the data processing module;

[0104] Data processing module: connected to the signal acquisition module, including a time-frequency analysis combined with a wavelet transform unit and a data enhancement unit. The time-frequency analysis combined with a wavelet transform unit processes the received radar echo signal into three-dimensional time-frequency-wavelet coefficient space data. The data enhancement unit rotates, flips, adds random noise to the three-dimensional data, and simulates the motion trajectories of different targets.

[0105] Feature extraction module: connected to the data processing module, including a three-dimensional convolutional neural network unit and a long short-term memory network unit. The 3D-CNN unit has the function of adaptively adjusting the size of the convolution kernel according to the complexity of the signal in different time-frequency-wavelet coefficient regions, and is used to extract the local features of three-dimensional data in the spatial dimension. The LSTM unit is used to capture the dynamic change characteristics of the signal in the time dimension;

[0106] Feature fusion and target detection module: connected to the feature extraction module, including a feature fusion unit based on a multimodal attention mechanism and a target detection unit based on a generative adversarial network. The feature fusion unit dynamically allocates weights according to the importance of different feature modalities to target detection to fuse the features extracted by 3D-CNN and LSTM. The target detection unit detects targets by generating possible target signal hypotheses through a generator and judging the similarity between the hypotheses and the real target signals through a discriminator.

[0107] Model training and optimization module: connected to the feature fusion and target detection module, including an adaptive learning rate adjustment unit and an early stopping control unit. The adaptive learning rate adjustment unit dynamically adjusts the learning rate according to the loss change of the model on the validation set, and the early stopping control unit stops training when the performance index of the model on the validation set does not improve for a specific training cycle;

[0108] Display module: connected to the model training and optimization module, used to receive the judgment results of the model output on the existence of the target, as well as the target's distance, speed, amplitude and target movement trend characteristic information, and visualize the target information. For faint targets, a special highlight display method is used. The highlight display method includes a flashing effect and an increase in the target mark size. The flashing frequency and the mark size increase ratio are adaptively adjusted according to the faintness of the target. The faintness is measured by the target signal-to-noise ratio.

[0109] The above specific embodiments are only several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A radar weak target detection method based on deep learning, characterized in that: The method comprises the following steps: Data preprocessing and enhancement: Receive the original radar echo signal, use time-frequency analysis combined with wavelet transform to generate three-dimensional time-frequency-wavelet coefficient space data, and expand the data set in various ways; Feature extraction network construction: Build a network that integrates 3D-CNN and LSTM. 3D-CNN has convolution operations with variable convolution kernel sizes. Feature fusion and target detection: Use a module based on a multimodal attention mechanism to fuse features, and input the fused features into a target detection module based on GAN; Model training and optimization: Train the network with labeled datasets, using adaptive learning rate adjustment algorithm and early stopping method; Real-time detection and result output: The real-time echo signal is pre-processed and input into the trained model, and the target information is output and visualized, with special display for weak targets.

2. The radar weak target detection method based on deep learning according to claim 1 is characterized in that: The data processing and enhancement specifically includes: receiving the original radar echo signal, processing the signal by using time-frequency analysis combined with wavelet transform, and generating three-dimensional time-frequency-wavelet coefficient space data, wherein the time-frequency analysis is used to convert the time domain signal into the time-frequency domain representation, and the wavelet transform is used to further highlight the characteristics of the signal at different scales; The generated 3D data is enhanced by rotating, flipping, adding random noise, and simulating different target motion trajectories to expand the data set so that the data set contains more types of weak target signals and noise mixtures. The specific formula of the time-frequency analysis combined with wavelet transform is: Assuming the radar echo signal is s(τ), its short-time Fourier transform STFT(t,f) is expressed as: Where ω(τ) is the window function; Wavelet transform WT(a,b) is expressed as: Among them, a is the scale factor, b is the translation factor, and ψ(t) is the wavelet basis function; the combined three-dimensional data is expressed as D(t,f,a)=STFT(t,f)×WT(a,b), where b is dynamically determined according to the signal characteristics and the target motion state.

3. The radar weak target detection method based on deep learning according to claim 1 is characterized in that: The feature extraction network construction specifically includes the following steps: Construct a feature extraction network that integrates a three-dimensional convolutional neural network and a long short-term memory network, where 3D-CNN is used to extract local features of three-dimensional data in the spatial dimension, and LSTM is used to capture the dynamic change characteristics of the signal in the temporal dimension; In the convolution layer of 3D-CNN, a convolution operation with a variable convolution kernel size is designed to adaptively adjust the convolution kernel size according to the complexity of the signal in different time-frequency-wavelet coefficient regions; The convolution operation with variable convolution kernel size is performed at the convolution kernel size K of the i-th convolution layer. i The calculation formula is: Among them, K0 is the basic convolution kernel size, α is the adjustment coefficient, M is the number of channels of the input feature map, and c i-1 (j) is the feature map of the jth channel in the previous layer.

4. The radar weak target detection method based on deep learning according to claim 1 is characterized in that: The feature fusion and target detection specifically include: After the feature extraction network, a feature fusion module based on the multimodal attention mechanism is designed. This module can dynamically assign weights according to the importance of different feature modalities to target detection and fuse the features extracted by 3D-CNN and LSTM. The fused features are input into a target detection module based on a generative adversarial network (GAN). The generator is used to generate possible target signal hypotheses, and the discriminator is used to judge the similarity between the generated hypotheses and the real target signals. The accuracy of target detection is improved through adversarial training. In the feature fusion module based on the multimodal attention mechanism, the attention weight w mn The calculation formula is: Among them, e mn It is the associated energy value of the mth 3D-CNN feature and the nth LSTM feature calculated by a bilinear mapping function. P and Q are the number of 3D-CNN and LSTM output features respectively. The calculation formula of the fused feature F is: Among them, F 3D-CNN (m) is the mth feature output by 3D-CNN, F LSTM (n) is the nth feature output by LSTM.

5. The radar weak target detection method based on deep learning according to claim 1 is characterized in that: The model training and optimization specifically includes the following steps: The entire network is trained using a large number of annotated data sets containing weak target signals. During the training process, an adaptive learning rate adjustment algorithm is used to dynamically adjust the learning rate according to the loss changes of the model on the validation set. Use the early stopping method to prevent overfitting. When the performance indicators of the model on the validation set do not improve for several consecutive training cycles, stop training and save the current optimal model. The adaptive learning rate adjustment algorithm is: assuming that the current learning rate is η, the training cycle is t, and the validation set loss is L(t), then the learning rate η(t+1) of the next training cycle is calculated as follows: Among them, β is the learning rate adjustment factor, when η(t+1)>η max When η(t+1)=η max , when η(t+1)<η min When η(t+1)=η min , η max and η min are the upper and lower limits of the learning rate.

6. The radar weak target detection method based on deep learning according to claim 5 is characterized in that: The number of consecutive training cycles in the early stopping method is N, which is dynamically determined according to the model complexity and the size of the data set, and the value range is between 5-15.

7. The radar weak target detection method based on deep learning according to claim 1, characterized in that: The real-time detection and result output are specifically as follows: The radar echo signal received in real time is input into the trained model after data preprocessing. The model outputs the judgment result of whether the target exists and the characteristic information of the target, including distance, speed, amplitude and target movement trend; The output results are visualized and the target information is intuitively displayed on the radar screen through different colors and marks. A special highlighting method is used for weak targets to facilitate identification.

8. The radar weak target detection method based on deep learning according to claim 7 is characterized in that: The special highlighting mode is to use a flickering effect and increase the target mark size. The flickering frequency and the mark size increase ratio are adaptively adjusted according to the target's faintness, and the faintness is measured by the target signal-to-noise ratio.

9. The radar weak target detection method based on deep learning according to claim 8, characterized in that: The labeled data set containing weak target signals includes the actual existence of the target, the precise distance, speed, amplitude of the target, and the starting point and end point of the target's motion trajectory. The labeling is completed by combining high-precision radar measurement equipment with manual auxiliary labeling.

10. A radar weak target detection device based on deep learning, using the radar weak target detection method based on deep learning as described in any one of claims 1 to 9, characterized in that: include: Signal acquisition module: used to receive radar echo signals and transmit them to the data processing module; Data processing module: connected to the signal acquisition module, including a time-frequency analysis combined with a wavelet transform unit and a data enhancement unit. The time-frequency analysis combined with a wavelet transform unit processes the received radar echo signal into three-dimensional time-frequency-wavelet coefficient space data. The data enhancement unit rotates, flips, adds random noise to the three-dimensional data, and simulates the motion trajectories of different targets. Feature extraction module: connected to the data processing module, including a three-dimensional convolutional neural network unit and a long short-term memory network unit. The 3D-CNN unit has the function of adaptively adjusting the size of the convolution kernel according to the complexity of the signal in different time-frequency-wavelet coefficient regions, and is used to extract the local features of three-dimensional data in the spatial dimension. The LSTM unit is used to capture the dynamic change characteristics of the signal in the time dimension; Feature fusion and target detection module: connected to the feature extraction module, including a feature fusion unit based on a multimodal attention mechanism and a target detection unit based on a generative adversarial network. The feature fusion unit dynamically allocates weights according to the importance of different feature modalities to target detection to fuse the features extracted by 3D-CNN and LSTM. The target detection unit detects targets by generating possible target signal hypotheses through a generator and judging the similarity between the hypotheses and the real target signals through a discriminator. Model training and optimization module: connected to the feature fusion and target detection module, including an adaptive learning rate adjustment unit and an early stopping control unit. The adaptive learning rate adjustment unit dynamically adjusts the learning rate according to the loss change of the model on the validation set, and the early stopping control unit stops training when the performance index of the model on the validation set does not improve for a specific training cycle; Display module: connected to the model training and optimization module, used to receive the judgment result of the model output on the existence of the target and the target's distance, speed, amplitude and target movement trend characteristic information, and visualize the target information. For faint targets, a special highlight display method is used. The highlight display method includes a flashing effect and an increase in the target mark size. The flashing frequency and the mark size increase ratio are adaptively adjusted according to the faintness of the target. The faintness is measured by the target signal-to-noise ratio.

Images