A target radar signal enhancement method and device based on DPRNN

Abstract

This application discloses a target radar signal enhancement method and apparatus based on DPRNN. First, a dataset is constructed. The mixed signals from the training set in the dataset are processed using sliding window segmentation to obtain mixed time-domain signals. Second, a target radar signal enhancement model is constructed based on the DPRNN network. The mixed time-domain signals are then enhanced to obtain reconstructed time-domain signals. The reconstructed time-domain signals are then input into the target radar signal enhancement model, and the model is trained using the training set in the dataset to obtain a trained target radar signal enhancement model. Finally, the trained target radar signal enhancement model is used to process the online-acquired mixed time-domain signals to obtain the enhanced target time-domain signal. By constructing a target radar signal enhancement model to extract contextual features of radar signals, the problem of target signal enhancement under time-frequency overlap conditions is solved, improving the signal enhancement effect.

Description

Technical Field

[0001] This application relates to the field of signal enhancement technology, and more specifically, to a target radar signal enhancement method and apparatus based on DPRNN. Background Technology

[0002] In the field of signal processing, especially for enhancing low signal-to-noise ratio (SNR) signals in complex environments, both traditional signal processing methods and emerging deep learning technologies have certain limitations. Traditional signal enhancement methods mainly include wavelet transform (WT), empirical mode decomposition (EMD) based on Hilbert-Huang transform (HHT), singular spectrum analysis (SSA), and independent component analysis (ICA).

[0003] While traditional methods have achieved some success in specific application scenarios, they perform poorly in low signal-to-noise ratio environments. Furthermore, due to a lack of prior information, they struggle to accurately distinguish target signals from background noise in complex environments, easily leading to the loss of critical pulse signals. With the development of deep learning technology, some research has begun to explore its application in signal enhancement. However, current deep learning-based signal enhancement schemes still face several challenges: on the one hand, these models often ignore contextual information in radar signals, limiting their generalization ability in real-world environments; on the other hand, when faced with overlapping target and background signals in the time-frequency domain, existing deep learning methods rely on the assumption of time-frequency independence, making it difficult to effectively separate the two and easily causing the loss of target signal pulses.

[0004] Therefore, there is an urgent need for a new method that can overcome the shortcomings of existing technologies. This method should not only maintain good robustness under low signal-to-noise ratio conditions, but also have the ability to accurately extract and enhance the target signal from complex background environments. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of existing technologies and provide a target radar signal enhancement method and apparatus based on DPRNN. By constructing a target radar signal enhancement model to extract the contextual features of the radar signal, the method solves the target signal enhancement problem under time-frequency overlap conditions and improves the signal enhancement effect.

[0006] The objective of this application is achieved through the following technical solution:

[0007] In a first aspect, this application proposes a target radar signal enhancement method based on DPRNN, the method comprising:

[0008] Step S1: Construct a dataset and perform sliding window segmentation on the mixed signals of the training set in the dataset to obtain mixed time-domain signals;

[0009] Step S2: Construct a target radar signal enhancement model based on the DPRNN network, and perform signal enhancement processing on the mixed time-domain signal to obtain the reconstructed time-domain signal;

[0010] Step S3: Input the reconstructed time-domain signal into the target radar signal enhancement model, and use the training set in the dataset to train the target radar signal enhancement model to obtain the trained target radar signal enhancement model.

[0011] Step S4: Use the trained target radar signal enhancement model to process the online intercepted hybrid time-domain signal to obtain the enhanced target time-domain signal.

[0012] In one possible implementation, step S1 includes:

[0013] Step S11: Obtain multiple original mixed signals and corresponding target radar signal time-domain sample pairs to construct a dataset;

[0014] Step S12: Divide the dataset into training set, validation set and test set according to the proportions;

[0015] Step S13: Perform sliding window segmentation on the mixed signal in the training set to obtain the mixed time domain signal.

[0016] In one possible implementation, the target radar signal enhancement model includes an encoder, a masking estimation module, and a decoder, and step S2 includes:

[0017] Step S21: Construct an encoder using a one-dimensional convolutional layer, and map the mixed temporal signals into a high-dimensional feature tensor through the encoder;

[0018] Step S22: Extract contextual information of high-dimensional feature tensors using the masking estimation module to generate high-dimensional spatial masks;

[0019] Step S23: Multiply the high-dimensional space mask and the high-dimensional feature tensor point by point to obtain the enhanced high-dimensional representation;

[0020] Step S24: Construct a decoder using one-dimensional transposed convolution, and reconstruct the original one-dimensional time-domain signal waveform from the enhanced high-dimensional representation using the decoder to obtain the reconstructed time-domain signal.

[0021] In one possible implementation, the mapping rule in step S21 is: w = H(conv1D(x)k )), where conv1D is a one-dimensional convolutional filter, H(·) is the PReLU activation function, and x k w represents the small blocks of the mixed time-domain signal, where w is the value of x. k The n-dimensional representation of the transformation.

[0022] In one possible implementation, the masking estimation module includes a normalization layer, a convolutional layer, a DPRNN network, a gated linear unit, and a rectified linear unit. Step S22 includes:

[0023] Step S22-1: After normalizing the high-dimensional feature tensor using a normalization layer, the normalized high-dimensional feature tensor is obtained.

[0024] Step S22-2: Use convolutional layers to adjust the number of channels or feature dimensions of the normalized high-dimensional feature tensor to obtain the processed high-dimensional feature tensor;

[0025] Step S22-3: Use the DPRNN network to extract context information from the processed high-dimensional feature tensor to generate a high-dimensional spatial mask.

[0026] In one possible implementation, steps S22-3 include:

[0027] The processed high-dimensional feature tensor is segmented by a sliding window to obtain multiple overlapping region blocks, which are then filled with zeros at both ends. Finally, all overlapping region blocks are spliced ​​together to form a three-dimensional tensor.

[0028] Based on three-dimensional tensors, long short-term memory networks are used to perform intra-block and inter-block processing to extract the contextual information of the signal.

[0029] The overlapping regions of contextual information are added together to restore the original dimension and generate a high-dimensional mask.

[0030] In one possible implementation, step S3 includes:

[0031] Initialize the target radar signal enhancement model and input the reconstructed time-domain signal into the initialized target radar signal enhancement model;

[0032] Input mixed signal samples from the training set, based on the constructed loss function SI-SNR Loss Loss is calculated through forward propagation;

[0033] The model parameters are updated iteratively through backpropagation. When the loss does not decrease on the validation set for ten consecutive rounds, training is stopped and the optimal model parameters are saved to obtain the trained target radar signal enhancement model.

[0034] In one possible implementation, the loss function is SI-SNR.Loss for: X E =X * -X T , where X * Indicates network output. This represents the actual target signal.

[0035] Secondly, this application proposes a target radar signal enhancement device based on DPRNN, the device comprising:

[0036] The building module is used to build the dataset and perform sliding window segmentation on the mixed signals of the training set in the dataset to obtain the mixed time-domain signals;

[0037] The signal enhancement module is used to construct a target radar signal enhancement model based on the DPRNN network and perform signal enhancement processing on the mixed time-domain signal to obtain the reconstructed time-domain signal.

[0038] The training module is used to input the reconstructed time-domain signal into the target radar signal enhancement model, and to train the target radar signal enhancement model using the training set in the dataset to obtain the trained target radar signal enhancement model.

[0039] The generation module is used to process the online-acquired hybrid time-domain signal using the trained target radar signal enhancement model to obtain the enhanced target time-domain signal.

[0040] In one possible implementation, the building module is used for:

[0041] A dataset is constructed by obtaining multiple pairs of original mixed signals and corresponding target radar signals in the time domain.

[0042] The dataset is divided into training, validation, and test sets according to a set ratio;

[0043] The mixed signals in the training set are processed by sliding window segmentation to obtain the mixed time-domain signals.

[0044] The main solution and its various further alternatives described above can be freely combined to form multiple solutions, all of which are solutions that can be adopted and are claimed in this application; furthermore, the (non-conflicting alternatives) can also be freely combined with each other and with other alternatives. Those skilled in the art, after understanding the solution of this application, will realize from the prior art and common general knowledge that there are many combinations, all of which are technical solutions to be protected by this application, and will not be exhaustively listed here.

[0045] This application discloses a target radar signal enhancement method and apparatus based on DPRNN. First, a dataset is constructed. The mixed signals from the training set in the dataset are processed using sliding window segmentation to obtain mixed time-domain signals. Second, a target radar signal enhancement model is constructed based on the DPRNN network. The mixed time-domain signals are then enhanced to obtain reconstructed time-domain signals. The reconstructed time-domain signals are then input into the target radar signal enhancement model, and the model is trained using the training set in the dataset to obtain a trained target radar signal enhancement model. Finally, the trained target radar signal enhancement model is used to process the online-acquired mixed time-domain signals to obtain the enhanced target time-domain signal. By constructing a target radar signal enhancement model to extract contextual features of radar signals, the problem of target signal enhancement under time-frequency overlap conditions is solved, improving the signal enhancement effect. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 A flowchart illustrating a target radar signal enhancement method based on DPRNN proposed in an embodiment of this application is shown.

[0048] Figure 2 A schematic diagram of the target radar signal enhancement model proposed in the embodiments of this application is shown.

[0049] Figure 3 A schematic diagram of the DPRNN network proposed in an embodiment of this application is shown.

[0050] Figure 4a This is a schematic diagram of the target signal proposed in the embodiments of this application.

[0051] Figure 4b This is a schematic diagram of the mixed signal before suppression as proposed in the embodiments of this application.

[0052] Figure 4c This is a schematic diagram of the enhanced signal under low signal-to-noise ratio and time-frequency overlap conditions proposed in the embodiments of this application.

[0053] Figure 5a This is a time-frequency diagram of the target signal proposed in the embodiments of this application.

[0054] Figure 5b This is a time-frequency diagram of the mixed signal proposed in the embodiments of this application.

[0055] Figure 5c This is a time-frequency diagram of the target signal enhancement result proposed in the embodiments of this application. Detailed Implementation

[0056] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0057] Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0058] In existing technologies, traditional signal enhancement methods are sensitive to noise and have difficulty adapting to low signal-to-noise ratio environments. Existing deep learning methods lack effective utilization of signal context information and rely on the independence of signal time and frequency. In cases of time-frequency overlap, it is difficult to distinguish between target signals and background signals, which can easily lead to the loss of target signal pulses.

[0059] Therefore, to address the aforementioned technical problems, this application proposes a target radar signal enhancement method and apparatus based on DPRNN. Starting from mid-frequency long-sequence data, a set of convolutional encoders and decoders is used to obtain a high-dimensional optimized representation of the one-dimensional signal. Simultaneously, stacked DPRNN networks are used to enhance the model's long-sequence modeling capability and extract contextual information of the signal to distinguish between target and background signals. Through the application of deep learning techniques, particularly in convolutional and recurrent networks, the recognition and enhancement capabilities of radar target signals are significantly improved, which will be described in detail below.

[0060] Please refer to Figure 1 , Figure 1 This paper presents a flowchart illustrating a target radar signal enhancement method based on DPRNN proposed in an embodiment of this application. The method includes:

[0061] Step S1: Construct a dataset and perform sliding window segmentation on the mixed signals of the training set in the dataset to obtain mixed time-domain signals.

[0062] Multiple original mixed signals and corresponding target radar signal time-domain sample pairs were obtained to construct an initial dataset. The dataset was then divided into a training set, a validation set, and a test set in a 6:2:2 ratio. The mixed signals in the training set were processed by sliding window segmentation to obtain mixed time-domain signal segments suitable for network training.

[0063] Step S1 includes:

[0064] Step S11: Obtain multiple pairs of original mixed signals and corresponding target radar signal time-domain samples to construct a dataset.

[0065] To collect a sufficient number of raw mixed signals and their corresponding target radar signal samples to construct a high-quality dataset, the following steps are taken: First, raw mixed signal samples containing background noise and other interference signals are acquired. These signals can be derived from the actual environment or generated through simulation. Then, target radar signal samples corresponding to each mixed signal are simultaneously acquired. The target signals are the clean radar signals that are to be extracted from the mixed signals. Each pair of raw mixed signals and its corresponding target radar signal is combined into a sample pair. All sample pairs constitute the original dataset. Where D is the original set of mixed signal samples. It is the corresponding set of target radar signal samples.

[0066] Step S12: Divide the dataset into training set, validation set and test set according to the proportion.

[0067] The dataset was divided into a training set and a DataSet in a 6:2:2 ratio. train Validation set DataSet val and test set DataSet test The dataset is divided into three parts: a training set (60%) for model training, a validation set (20%) for adjusting model parameters and preventing overfitting, and a test set (20%) for final model performance evaluation. This division also ensures the dataset is randomly partitioned to avoid bias and guarantee the representativeness of each subset. By rationally partitioning the dataset, sufficient representative samples are ensured for the model training, validation, and testing phases.

[0068] Step S13: Perform sliding window segmentation on the mixed signal in the training set to obtain the mixed time domain signal.

[0069] Choosing 20 milliseconds as the sliding window size and setting an appropriate overlap ratio (e.g., 50%) as needed to ensure sufficient information overlap between adjacent segments, each mixed signal is segmented according to a 20ms time window to obtain multiple time-domain segments. Each segment serves as an independent input sample, and each segmented time-domain segment is used as input for network training, forming a new training set. The mixed signal time-domain signal samples in the training set are then cut into fixed-length segments as input for network training.

[0070] Step S2: Construct a target radar signal enhancement model based on the DPRNN network, and perform signal enhancement processing on the mixed time domain signal to obtain the reconstructed time domain signal.

[0071] The target radar signal enhancement model based on the DPRNN network converts the mixed time-domain signal into a high-dimensional feature representation through a one-dimensional convolutional layer. Then, the DPRNN is used to extract contextual information to generate a masking matrix to suppress noise and preserve the target signal. Finally, a clear time-domain signal is reconstructed through a transposed convolutional layer, thereby effectively improving the quality of the radar signal.

[0072] Please refer to Figure 2 , Figure 2 The diagram illustrates a target radar signal enhancement model proposed in this application. The model includes an encoder, a masking estimation module, and a decoder. The encoder processes the input signal to generate a masking estimate, and the masking estimate output by the encoder is used to enhance the input signal, generating an enhanced signal. The decoder further processes the enhanced signal to generate the final output signal. In the masking estimation module, 1×Conv represents a convolutional layer that performs initial processing on the input signal. A deep recurrent neural network is used to process and enhance the signal. A gated linear unit (GLU) is used to control the information flow, and ReLU is the activation function used to introduce nonlinear characteristics.

[0073] Step S2 includes:

[0074] Step S21: Construct an encoder using a one-dimensional convolutional layer to map the mixed temporal signals into a high-dimensional feature tensor.

[0075] An encoder is constructed using one-dimensional convolutional blocks to map the input one-dimensional signal x into a high-dimensional feature tensor. The convolutional layer divides x into several overlapping small blocks. (kernel_size represents the filter length), and is converted to an n-dimensional representation (number of filters). In step S21, the mapping rule is: w = H(conv1D(x) k )), where conv1D is a one-dimensional convolutional filter, H(·) is the PReLU activation function, and x k w represents the small blocks of the mixed time-domain signal, where w is the value of x. k The n-dimensional representation of the transformation.

[0076] Step S22: Use the masking estimation module to extract the contextual information of the high-dimensional feature tensor and generate a high-dimensional spatial mask.

[0077] The masking estimation module extracts contextual information from high-dimensional feature tensors through the DPRNN network to generate a high-dimensional spatial masking matrix. This masking matrix is ​​used to identify and enhance target signal components while suppressing background noise and other interference, thereby improving signal quality.

[0078] The masking estimation module includes a normalization layer, a convolutional layer, a DPRNN network, gated linear units, and rectified linear units. Step S22 includes:

[0079] Step S22-1: After normalizing the high-dimensional feature tensor using a normalization layer, the normalized high-dimensional feature tensor is obtained.

[0080] The input high-dimensional feature tensor is normalized using a normalization layer to ensure that its numerical distribution is consistent, resulting in a normalized high-dimensional feature tensor.

[0081] Step S22-2: Use convolutional layers to adjust the number of channels or feature dimensions of the normalized high-dimensional feature tensor to obtain the processed high-dimensional feature tensor;

[0082] The number of channels or feature dimension of the normalized high-dimensional feature tensor is adjusted by using a 1×1 convolutional layer to make it suitable for input into the subsequent DPRNN network, thus obtaining the processed high-dimensional feature tensor.

[0083] Step S22-3: Use the DPRNN network to extract context information from the processed high-dimensional feature tensor to generate a high-dimensional spatial mask.

[0084] The processed high-dimensional feature tensor is input into a DPRNN (Dual-Path RNN) network. In the DPRNN network, gated linear units (GLUs) are used to preserve temporal information and improve the accuracy of output masking. Rectified linear units (ReLUs) are used to ensure the non-negativity of output masking, and finally, a high-dimensional spatial mask is generated.

[0085] Step S22-3 includes:

[0086] The processed high-dimensional feature tensor is segmented by a sliding window to obtain multiple overlapping region blocks, which are then filled with zeros at both ends. Finally, all overlapping region blocks are spliced ​​together to form a three-dimensional tensor.

[0087] Based on three-dimensional tensors, long short-term memory networks are used to perform intra-block and inter-block processing to extract the contextual information of the signal.

[0088] The overlapping regions of contextual information are added together to restore the original dimension and generate a high-dimensional mask.

[0089] Figure 3 This diagram illustrates the DPRNN network proposed in an embodiment of this application. The DPRNN consists of three parts: block segmentation, block processing, and overlapping summation. The input is divided into B overlapping blocks of length P using a sliding window of size Win. The first and last blocks are padded with zeros to ensure all blocks are of equal size. All blocks are then concatenated to form a three-dimensional tensor. By using stacked long short-term memory networks for intra-block and inter-block processing, signal context information is obtained; the outputs are then overlapped and added together to restore the input dimension.

[0090] Step S23: Multiply the high-dimensional space mask and the high-dimensional feature tensor point by point to obtain the enhanced high-dimensional representation.

[0091] The generated high-dimensional mask M (where M represents the energy proportion of the target signal in X) is multiplied pointwise with the original high-dimensional feature tensor to eliminate other components, retaining only the energy proportion of the target signal, thus obtaining the enhanced high-dimensional representation of the target radar signal. This enables effective enhancement of mixed time-domain signals.

[0092] Step S24: Construct a decoder using one-dimensional transposed convolution, and reconstruct the original one-dimensional time-domain signal waveform from the enhanced high-dimensional representation using the decoder to obtain the reconstructed time-domain signal.

[0093] The decoder module is constructed using one-dimensional transposed convolutions, ensuring that the transposed convolutions in the decoder have the same stride and kernel size as the convolutions in the encoder, thus guaranteeing that the size of the reconstructed signal is consistent with the original input. Then, the high-dimensional representation of the target radar signal obtained after masking enhancement is input into the decoder module. Through a series of transposed convolution operations, the spatial dimension of the signal is gradually restored, ultimately generating the reconstructed time-domain waveform. After processing by the decoder, the reconstructed time-domain signal is output. This signal is as close as possible to the original input signal in the time domain, but noise and other interference have been effectively suppressed, resulting in a significant enhancement of the target signal.

[0094] Step S3: Input the reconstructed time-domain signal into the target radar signal enhancement model, and use the training set in the dataset to train the target radar signal enhancement model to obtain the trained target radar signal enhancement model.

[0095] The reconstructed time-domain signal is input into the target radar signal enhancement model, and the model is trained using the training set in the dataset to optimize the model parameters. During training, the model learns how to extract and enhance the target radar signal from the mixed signal, ultimately resulting in a fully trained and performance-optimized target radar signal enhancement model.

[0096] Step S3 includes:

[0097] Initialize the target radar signal enhancement model and input the reconstructed time-domain signal into the initialized target radar signal enhancement model;

[0098] Input mixed signal samples from the training set, based on the constructed loss function SI-SNR Loss Loss is calculated through forward propagation;

[0099] The model parameters are updated iteratively through backpropagation. When the loss does not decrease on the validation set for ten consecutive rounds, training is stopped and the optimal model parameters are saved to obtain the trained target radar signal enhancement model.

[0100] The parameters of the target radar signal enhancement model are randomly initialized. The reconstructed time-domain signal is input into the initialized target radar signal enhancement model as initial input data. Mixed signal samples from the training set are used as input to the model. Each sample contains background noise and other interference components, as well as the corresponding target radar signal. Through the forward propagation process of the model, the difference between the predicted output and the real target signal is calculated, and the constructed loss function is used to measure the difference.

[0101] Based on the calculated loss value, the gradient is calculated using the backpropagation algorithm, and the model parameters are iteratively updated using an optimization algorithm to minimize the loss function. During training, the model performance is periodically evaluated on the validation set. When the loss on the validation set no longer decreases after ten consecutive rounds, an "early stop" mechanism is triggered, stopping training and saving the current optimal model parameters. Finally, the best model parameters obtained after training and optimization are saved, thus obtaining a well-trained model that can effectively extract and enhance target radar signals.

[0102] Loss function SI-SNR Loss for: X E =X * -X T , where X * Indicates network output. This represents the actual target signal.

[0103] Step S4: Use the trained target radar signal enhancement model to process the online intercepted hybrid time-domain signal to obtain the enhanced target time-domain signal.

[0104] To segment the online-acquired mixed time-domain signal into small segments suitable for model input, the mixed time-domain signal acquired via online pipeline is segmented into 20-millisecond (ms) segments, ensuring that each segment meets the model input requirements. The segmented 20ms mixed signal segments are used as input to a trained and stored target radar signal enhancement model. The model performs forward propagation calculations and outputs enhanced target time-domain signal segments. Finally, the outputs of all segments are concatenated to reconstruct the complete enhanced time-domain signal.

[0105] By using a trained target radar signal enhancement model, the real-time intercepted mixed time-domain signal is processed, and the enhanced target time-domain signal is finally output, thereby improving the signal quality and resolvability.

[0106] In one possible embodiment, a target signal enhancement scenario is constructed, in which there is 1 target radar and 55 background radars. Table 1 shows the parameters of the transmitted signals of the target radar and background radars with random range settings:

[0107] Table 1

[0108]

[0109] Based on the constructed scene simulation data, 2000 samples were generated to build a training dataset.

[0110] Generate a corresponding number of target and background signals, randomly superimpose them in the time domain, add Gaussian white noise with a signal-to-noise ratio of -15 to 20 dB, generate mixed time-domain signal samples, and save the corresponding target signal samples. Generate 2000 such sample pairs to construct a training dataset. Where D is the original set of mixed signal samples; This is the set of corresponding target signal samples.

[0111] The dataset DataSet is divided into a training set and a training set in a 6:2:2 ratio. train Validation set DataSet val and test set DataSet test Their sizes are 1200, 400, and 400 respectively.

[0112] The mixed signal time-domain signal samples in the training set are processed by sliding window segmentation, cut into 20ms time-domain segments, and used as input for network training.

[0113] This paper implements a target signal enhancement model based on PyTorch, including an encoder, a masking estimation module, and a decoder.

[0114] An encoder is constructed using one-dimensional convolutional blocks, dividing x into several overlapping small blocks. And convert it to an n-dimensional representation w = H(conv1D(x) k In the formula, conv1D is a one-dimensional convolutional filtering process; kernel_size represents the filter length, which is 2 in this example; n represents the number of filters, which is 32 in this example; H(·) is the PReLU activation function.

[0115] A masking estimation module based on the DPRNN network is designed. The high-dimensional tensor X is input into the DPRNN network after passing through a normalization layer and a 1×1 convolutional layer. A gated linear unit (GLU) is used to preserve temporal information, improving the accuracy of the output masking. A rectified linear unit (ReLU) is used to ensure the non-negativity of the output masking. The network is constructed as follows: Figure 3 The DPRNN network of this method, as shown, consists of three parts: block division, block processing, and overlapping concatenation. The input is divided into B overlapping blocks of length P using a sliding window of size Win; in this example, Win = 400; B = 5002; P = 200. The first and last blocks are zero-padded to ensure all blocks are of equal size. All blocks are then concatenated to form a three-dimensional tensor. Signal context information is obtained by performing intra-block and inter-block processing using stacked Long Short-Term Memory (LSTM) networks. The outputs are then overlapped and summed to restore the input dimension. The masking estimation module generates a high-dimensional mask M; M represents the energy proportion of the target signal in X, which is multiplied point-by-point with X to eliminate other components, resulting in an enhanced high-dimensional representation of the target signal.

[0116] A decoder module is constructed using one-dimensional transposed convolution to reconstruct the temporal waveform from the masked and enhanced high-dimensional representation of the target:

[0117] Next, the loss function is calculated based on the network output and the target signal set in the dataset, specifically the SI-SNR loss function. Loss The model parameters are updated iteratively by minimizing the loss function.

[0118] For the target signal enhancement model, the mixed signal samples in the training set are used as the model input, and the target signal corresponding to the mixed samples is used as the target. SI-SNR is used as the target signal. Loss The model is trained using this loss function. Adam is chosen as the optimizer, with weights set to 1×10. -4 The initial learning rate is set to 1×10. -3 The maximum number of iterations is set to 100. During the training phase, the model parameters are first initialized, the mixed signal x is input, and the SI-SNR is calculated through forward propagation. Loss The loss is used to iteratively update the model parameters through backpropagation. When the loss on the validation set does not decrease for five consecutive epochs, the learning rate is adjusted to half of the current learning rate. An "early stop" training strategy is adopted: when the loss on the validation set does not decrease for ten consecutive epochs, training is stopped and the model parameters are saved.

[0119] For the mixed time-domain signal input, it is segmented into 20ms segments. For the sample x of the mixed signal segment, it is used as the input of the saved network model, and the enhanced target signal is output.

[0120] For mixed signals in a simulated scene, by using a target signal enhancement model to suppress background signals and noise, the following can be obtained: Figure 4c The diagram shown illustrates the enhanced signal under conditions of low signal-to-noise ratio and time-frequency overlap. Figure 4a 4b is a schematic diagram of the target signal, and 4b is a schematic diagram of the mixed signal before suppression.

[0121] Figure 5a This is the time-frequency diagram of the corresponding target signal. Figure 5b This is the time-frequency plot of the mixed signal. Figure 5c The time-frequency plot of the target signal enhancement result shows that this application can achieve target signal enhancement under conditions of low signal-to-noise ratio and time-frequency overlap. Table 2 compares the performance of the proposed signal enhancement method with other signal enhancement methods:

[0122] Table 2

[0123] SI-SNR (dB) This article's model 18.66 SSA-ICA 8.06 EMD-ICA -5.37

[0124] As can be seen, the method proposed in this application achieves a significant improvement in signal-to-noise ratio, and its signal enhancement performance has a significant advantage over other methods.

[0125] Compared with the prior art, the embodiments of this application have the following beneficial effects:

[0126] First, this application uses a one-dimensional convolutional codec instead of time-frequency transformation, which can more effectively extract and process time-domain features, thereby significantly improving the signal enhancement effect in low signal-to-noise ratio environments.

[0127] Second, by introducing a dual-path structure, local and global contextual information can be processed simultaneously. One path focuses on extracting short-term local features, while the other path focuses on learning long-term dependencies. This enables the model to better separate the target signal from background noise in complex time-frequency overlap situations, thereby achieving more accurate target signal enhancement.

[0128] Third, it performs well under conditions of low signal-to-noise ratio and time-frequency overlap, making up for the shortcomings of existing signal enhancement methods in these complex scenarios.

[0129] Fourth, the one-dimensional convolutional encoder-decoder and dual-path recurrent network structure adopted are more compact and efficient, which not only reduces the demand for computing resources, but also speeds up inference, making it suitable for real-time signal processing applications.

[0130] The following provides a possible implementation of a DPRNN-based target radar signal enhancement device, which performs the various execution steps and corresponding technical effects of the DPRNN-based target radar signal enhancement method shown in the above embodiments and possible implementations. The device includes:

[0131] The building module is used to build the dataset and perform sliding window segmentation on the mixed signals of the training set in the dataset to obtain the mixed time-domain signals;

[0132] The signal enhancement module is used to construct a target radar signal enhancement model based on the DPRNN network and perform signal enhancement processing on the mixed time-domain signal to obtain the reconstructed time-domain signal.

[0133] The training module is used to input the reconstructed time-domain signal into the target radar signal enhancement model, and to train the target radar signal enhancement model using the training set in the dataset to obtain the trained target radar signal enhancement model.

[0134] The generation module is used to process the online-acquired hybrid time-domain signal using the trained target radar signal enhancement model to obtain the enhanced target time-domain signal.

[0135] In one possible embodiment, the building module is configured to:

[0136] A dataset is constructed by obtaining multiple pairs of original mixed signals and corresponding target radar signals in the time domain.

[0137] The dataset is divided into training, validation, and test sets according to a set ratio;

[0138] The mixed signals in the training set are processed by sliding window segmentation to obtain the mixed time-domain signals.

[0139] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A target radar signal enhancement method based on DPRNN, characterized in that, The method includes: Step S1: Construct a dataset and perform sliding window segmentation on the mixed signals of the training set in the dataset to obtain mixed time-domain signals; Step S2: Construct a target radar signal enhancement model based on the DPRNN network, and perform signal enhancement processing on the mixed time-domain signal to obtain the reconstructed time-domain signal; the target radar signal enhancement model includes an encoder, a masking estimation module, and a decoder. Step S2 includes: Step S21: Construct an encoder using a one-dimensional convolutional layer, and map the mixed temporal signals into a high-dimensional feature tensor through the encoder; Step S22: Extract contextual information of high-dimensional feature tensors using the masking estimation module to generate high-dimensional spatial masks; Step S23: Multiply the high-dimensional space mask and the high-dimensional feature tensor point by point to obtain the enhanced high-dimensional representation; Step S24: Construct a decoder using one-dimensional transposed convolution, and reconstruct the original one-dimensional time-domain signal waveform from the enhanced high-dimensional representation using the decoder to obtain the reconstructed time-domain signal. Step S3: Input the reconstructed time-domain signal into the target radar signal enhancement model, and use the centralized training set of the dataset to train the target radar signal enhancement model to obtain the trained target radar signal enhancement model. Step S4: Use the trained target radar signal enhancement model to process the online intercepted hybrid time-domain signal to obtain the enhanced target time-domain signal.

2. The target radar signal enhancement method as described in claim 1, characterized in that, Step S1 includes: Step S11: Obtain multiple original mixed signals and corresponding target radar signal time-domain sample pairs to construct a dataset; Step S12: Divide the dataset into training set, validation set and test set according to the proportions; Step S13: Perform sliding window segmentation on the mixed signal in the training set to obtain the mixed time domain signal.

3. The target radar signal enhancement method as described in claim 1, characterized in that, The mapping rule in step S21 is: ,in For one-dimensional convolution filtering, The PReLU activation function is used. To divide the mixed time-domain signal into small blocks, To Transformation Wei said.

4. The target radar signal enhancement method as described in claim 1, characterized in that, The masking estimation module includes a normalization layer, a convolutional layer, a DPRNN network, gated linear units, and rectified linear units. Step S22 includes: Step S22-1: After normalizing the high-dimensional feature tensor using a normalization layer, the normalized high-dimensional feature tensor is obtained. Step S22-2: Use convolutional layers to adjust the number of channels or feature dimensions of the normalized high-dimensional feature tensor to obtain the processed high-dimensional feature tensor; Step S22-3: Use the DPRNN network to extract context information from the processed high-dimensional feature tensor to generate a high-dimensional spatial mask.

5. The target radar signal enhancement method as described in claim 4, characterized in that, Step S22-3 includes: The processed high-dimensional feature tensor is segmented by a sliding window to obtain multiple overlapping regions, which are then filled with zeros at both ends. Finally, all overlapping regions are spliced ​​together to form a three-dimensional tensor. Based on three-dimensional tensors, long short-term memory networks are used to perform intra-block and inter-block processing to extract the contextual information of the signal. The overlapping regions of contextual information are added together to restore the original dimension and generate a high-dimensional mask.

6. The target radar signal enhancement method as described in claim 4, characterized in that, Step S3 includes: Initialize the target radar signal enhancement model and input the reconstructed time-domain signal into the initialized target radar signal enhancement model; Then, input the mixed signal samples from the training set, based on the constructed loss function. Loss is calculated through forward propagation; The model parameters are updated iteratively through backpropagation. When the loss does not decrease on the validation set for ten consecutive rounds, training is stopped and the optimal model parameters are saved to obtain the trained target radar signal enhancement model.

7. The target radar signal enhancement method as described in claim 6, characterized in that, loss function for: , , ,in Indicates network output. This represents the actual target signal.

8. A target radar signal enhancement device based on DPRNN, characterized in that, The device includes: The building module is used to build the dataset and perform sliding window segmentation on the mixed signals of the training set in the dataset to obtain the mixed time-domain signals; The signal enhancement module is used to build an encoder using a one-dimensional convolutional layer, which maps the mixed temporal signals into a high-dimensional feature tensor. The masking estimation module is used to extract the contextual information of the high-dimensional feature tensor to generate a high-dimensional spatial mask. The enhanced high-dimensional representation is obtained by multiplying the high-dimensional space masking with the high-dimensional feature tensor point by point. A decoder is constructed using one-dimensional transposed convolution. The original one-dimensional time-domain signal waveform is reconstructed from the enhanced high-dimensional representation using the decoder, and the reconstructed time-domain signal is obtained. The training module is used to input the reconstructed time-domain signal into the target radar signal enhancement model, and to train the target radar signal enhancement model using the training set in the dataset to obtain the trained target radar signal enhancement model. The generation module is used to process the online-acquired hybrid time-domain signal using the trained target radar signal enhancement model to obtain the enhanced target time-domain signal.

9. The target radar signal enhancement device as described in claim 8, characterized in that, Build modules are used for: A dataset is constructed by obtaining multiple pairs of original mixed signals and corresponding target radar signals in the time domain. The dataset is divided into training, validation, and test sets according to a certain ratio; The mixed signals in the training set are processed by sliding window segmentation to obtain the mixed time-domain signals.

Images