Microwave photon radar multi-signal reconstruction method based on deep learning

Abstract

The invention specifically discloses a microwave photon radar multi-signal reconstruction method based on deep learning, and relates to the technical field of information systems. The method comprises the following steps: S1, generating multi-target microwave radar data through simulation; s2, performing preprocessing and frequency domain signal extraction on the multi-target microwave radar data; s3, constructing a deep learning network model, and performing feature dimension reduction and feature extraction through a 1D-Unet encoder; s4, carrying out global feature extraction through a Transform network; s5, dimension raising is carried out on the processed low-dimensional features through a 1D-Unet decoder, and original signals are recovered; s6, combining time frequency processing and time domain reconstruction on the recovered original signal to generate a recovered multi-target signal; and S7, result testing and performance evaluation are carried out. The method significantly improves the recovery precision and robustness of multi-target signals in a complex electromagnetic environment, and is suitable for the fields of radar target recognition and tracking and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of information systems, and in particular to a microwave photon radar multi-signal reconstruction method based on deep learning. Background Art

[0002] Microwave photonics, a cutting-edge discipline that deeply integrates microwave and photonic technologies, leverages unique technical principles to modulate microwave signals into the optical domain for transmission and processing, demonstrating numerous outstanding advantages. Its high bandwidth meets the demands of massive data transmission; its extremely low transmission loss effectively reduces energy loss during signal transmission; its lightweight design supports lightweight system design; its rapid reconfigurability allows for flexible adaptation to diverse application scenarios; and its robust resistance to electromagnetic interference ensures the stability and reliability of signal transmission. These advantages make microwave photonics highly compatible with the development trend of radar, communications, and electronic warfare systems toward array-based, networked, integrated, and intelligent systems. It is widely used in numerous key areas, including national defense security, communications network construction, imaging technology, and precision instrument manufacturing.

[0003] In recent years, the cross-integration of microwave photonics and artificial intelligence has become an emerging and highly promising development direction in this field. With the help of artificial intelligence methods, it is expected that new ways of converting and manipulating microwaves and light waves will be explored, opening up new paths for improving the performance of related devices, thereby better meeting the growing high demands of microwave photonics technology in systems such as communications, radar, and electronic detection. Compared with traditional electronic technology, microwave photonics technology has significant advantages in signal processing. It not only provides high-frequency and multi-band processing capabilities, but also successfully breaks through the bandwidth limitations and high power consumption bottlenecks faced by traditional electronic technology in high-frequency signal processing. At the same time, microwave photonics fully utilizes the unique advantages of optical fiber transmission to achieve long-distance, high-speed, and low-loss signal transmission, and has strong anti-interference and dynamic adjustment capabilities, making it exhibit excellent adaptability in complex and changing application environments.

[0004] In practical applications, one of the key challenges facing microwave photonic technology is the need to accurately detect and recover multi-target signals in complex and diverse spectral environments, accurately extract the required information from complex original signals, and thus realize the core functions of radar systems such as target recognition, tracking, and environmental monitoring. Summary of the Invention

[0005] The purpose of this invention is to propose a microwave photonic radar multi-signal reconstruction method based on deep learning, aiming to solve the problems faced by the existing technology in multi-target signal recovery, such as spectrum interference, undersampling information loss, and insufficient long-term dependency modeling, and enhance the signal recovery capability of the radar system in complex electromagnetic environments.

[0006] To achieve the above objectives, the present invention proposes a microwave photon radar multi-signal reconstruction method based on deep learning, the steps of which are as follows:

[0007] Step S1, generating multi-target microwave radar data through simulation, specifically including multi-target signal generation, multi-channel notching and multi-channel undersampling extraction;

[0008] Step S2: preprocessing and frequency domain signal extraction of the multi-target microwave radar data, including data noise addition, data standardization and normalization, time-frequency analysis and label design, and dividing the data into a training set and a test set;

[0009] Step S3: construct a Unet-Transformer deep learning network model and perform feature dimensionality reduction and feature extraction through a 1D-Unet encoder;

[0010] Step S4: Capture the long-range dependency of the signal through the Transformer network and perform global feature extraction;

[0011] Step S5: Use the 1D-Unet decoder to increase the dimension of the low-dimensional features processed by the Transformer network and finally restore the original signal;

[0012] Step S6: signal time-frequency conversion, combining the restored original signal with time-frequency processing and time-domain reconstruction to generate a restored multi-target signal;

[0013] Step S7: Result testing and performance evaluation.

[0014] Preferably, in step S1, the multi-target signal generation includes four types of signals: Chirp, Modsinc, Filtered whitenoise and LFM; the multi-channel notch introduces different notch strategies for different channels, and the signal in the channel will pass through a specific filter or notch filter to simulate the interference between different targets or signal sources; the multi-channel undersampling extraction dynamically adjusts the number of sampling points according to the frequency characteristics and bandwidth requirements of different signals.

[0015] Preferably, in step S3, the encoder performs dimensionality reduction processing on the input original signal through multiple convolutional layers Conv1D. Each convolutional layer reduces the dimension of the signal and extracts features through pooling operations. Through multiple layers of convolution-pooling operations, the signal information is gradually compressed into a low-dimensional feature representation. The specific formula is as follows:

[0016] Assume the input signal is Where N is the length of the signal, D is the number of channels of the signal, and after multiple layers of convolution, a low-dimensional feature representation is obtained:

[0017] X encoded=Encoder(X);

[0018] Among them, Encoder(·) is the encoder function, which consists of multiple convolutional layers and pooling layers, X encoded is the low-dimensional output feature representation.

[0019] Preferably, in step S3, the Unet-Transformer deep learning network model uses the following loss function to optimize the entire network during training:

[0020] L total =L reconstruction +L regularization ;

[0021] Among them, L reconstruction To restore the signal to the original signal, L regularization is the regularization term, L total is the total loss function;

[0022] The mean square error is used as L reconstruction , the formula is as follows:

[0023]

[0024] Where N is the total number of samples, X i,true is the original signal sample, X i,reconstructed is the signal sample reconstructed by the model, and i is the sample index.

[0025] Preferably, in step S4, a Transformer network is used to capture the long-range dependency of the signal. Specifically, a multi-head self-attention mechanism is used to calculate the relationship between each time step in the signal and generate a new feature representation for each time step. The steps are as follows:

[0026] Q,K,V=Linear(X encoded );

[0027] Where Q is the query vector, K is the key vector, V is the value vector, and Linear(·) is the fully connected linear mapping function;

[0028] The attention of each position to all positions is calculated by scaling the dot product attention. The formula is as follows:

[0029]

[0030] Among them, d kis the dimension of query and key, T is the transposition operation, softmax(·) is the normalization function, and Attention(·) is the attention of each position to all positions;

[0031] The new feature representation is obtained through the feedforward neural network FFN. The formula is as follows:

[0032] X transformed =TransformerEncoder(X encoded );

[0033] Among them, X transformed is the encoded feature output, and TransformerEncoder(·) is the Transformer encoder module.

[0034] Preferably, in step S5, the decoder restores the feature map to the dimension of the original signal through a deconvolution layer and upsampling operation, including convolution, deconvolution and skip connection. The specific formula is as follows:

[0035] X reconstructed =Decoder(X transformed );

[0036] Among them, X reconstructed To restore the original signal, Decoder(·) is the decoder function, which consists of multiple deconvolution layers, upsampling layers and skip connections.

[0037] Preferably, in step S7, the specific steps of result testing and performance evaluation are as follows:

[0038] Step S71: Input the signals in the test set into the trained Unet-Transformer deep learning network model, and perform feature extraction, time-frequency analysis, long-range dependency capture, and signal recovery operations on each signal;

[0039] Step S72: Output the restored signal X reconstructed With the real signal X true Make a comparison;

[0040] Step S73: Use the evaluation indicators mean square error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM) to quantitatively analyze the quality of the restored signal.

[0041] Step S74: Display the comparison between the restored signal and the original signal through a time domain waveform diagram and a frequency domain spectrum diagram.

[0042] Therefore, the present invention proposes a microwave photon radar multi-signal reconstruction method based on deep learning, which has the following beneficial effects:

[0043] (1) Improving signal detection and recovery capabilities: By generating diverse and realistic training data through simulation and combining it with the Unet-Transformer deep learning network, this method can efficiently extract target signal features from complex spectrum data and accurately recover the signal spectrum. Compared with traditional methods, its detection and recovery capabilities for multiple target signals in complex electromagnetic environments are significantly improved, and different target signals can be more accurately identified and distinguished, providing a reliable data foundation for subsequent precise analysis of radar systems.

[0044] (2) Enhanced system adaptability and robustness: The introduction of deep learning technology significantly enhances the system's adaptability and robustness in the face of complex situations such as spectrum interference, noise, and signal obstruction. The Transformer network can effectively capture the long-range dependencies of signals, making up for the limitations of traditional convolutional neural networks, allowing the system to better understand the global information of signals. Even in a dynamically changing environment, it can stably and accurately process signals, ensuring the normal functioning of the radar system.

[0045] (3) Promote the widespread application of microwave photon radar: This method solves the key problem of signal processing of microwave photon radar in complex spectral environments, improves the system performance and reliability, and makes the advantages of microwave photon radar in practical applications more prominent. It lays a solid foundation for its widespread application in many fields such as national defense, communication networks, imaging and instruments, and is expected to promote the upgrading and development of related industry technologies.

[0046] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of a microwave photon radar multi-signal reconstruction method based on deep learning of the present invention;

[0048] Figure 2 Schematic diagram of the structure of the Unet-Transformer deep learning network model. DETAILED DESCRIPTION

[0049] To make the technical solutions, advantages, and objectives of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are part of the embodiments of the present invention, not all of them. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0051] like Figure 1 FIG. 1 is a flowchart of a microwave photon radar multi-signal reconstruction method based on deep learning according to the present invention. The specific steps are as follows:

[0052] S1. Multi-target microwave radar simulation data generation: Generate multi-target radar data through simulation, which is divided into three parts: multi-target signal generation, multi-channel notch and multi-channel notch undersampling extraction;

[0053] S11, Multi-target signal generation, aims to generate various types of signals with different parameters; specifically, it generates four types of analog signals, including: Chirp, Modsinc, Filtered white noise and LFM.

[0054] S12, multi-channel notching. After generating multi-target signals, the signals need to be processed through multiple channels. Different notching strategies are introduced for different channels to simulate interference in complex environments. The signal in each channel passes through a specific filter or notch to simulate interference between different targets or signal sources. The signal is filtered in multiple bands using a multi-channel notch function, with the notch center frequency and bandwidth set and stored as label data.

[0055] S13, Multi-channel Undersampling Decimation: To simulate real-world undersampling, the signal sampling rate is reduced through decimation. During undersampling, the number of sampling points is dynamically adjusted based on the frequency characteristics and bandwidth requirements of different signals. Undersampling results in spectrum loss, making signal recovery more difficult. Multi-channel undersampling decimation provides undersampled data for subsequent signal recovery algorithms, serving as input for deep learning models.

[0056] S2. Preprocess the multi-target microwave radar data and extract frequency domain signals, including data noise addition, data standardization and normalization, time-frequency analysis, and label design. The data is then divided into training and test sets. Specifically:

[0057] S21, data noise is added to simulate signal interference in real environment;

[0058] S22. Data standardization and normalization improve the training efficiency and convergence speed of deep learning models. Standardization can make each eigenvalue of the signal have the same scale by subtracting the mean and dividing by the standard deviation. Normalization usually adjusts the amplitude of the signal to a specific range, such as [0, 1] or [-1, 1].

[0059] S23, using short-time Fourier transform (STFT) to perform time-frequency analysis and extract the time-frequency characteristics of the signal;

[0060] S24. In the signal recovery process, label generation is crucial. In order to accurately add appropriate labels to each simulation signal, it is first necessary to clarify the basic characteristics of the signal, such as signal type, parameter settings, frequency range, time domain characteristics, etc.

[0061] S25. The data is divided into training set and test set. The specific operations are as follows:

[0062] S251. Maintain data representativeness: Training data and test data should come from similar distributions, that is, they have similar characteristics in terms of signal type, signal parameters, noise level, etc.

[0063] S252. Avoid data leakage: There should be no overlap between training data and test data, especially in the case of multiple target signals. The same target signal should be completely separated in training and testing.

[0064] S253. Ensure diversity in the test set: The test data should contain a variety of signal types and noise levels to simulate different scenarios in real applications and ensure the generalization ability of the model;

[0065] S254. Divide the dataset into a training set and a test set in a ratio of 70%:30%.

[0066] S3. Build a Unet-Transformer deep learning network model and perform feature dimensionality reduction and feature extraction through a 1D-Unet encoder.

[0067] like Figure 2 As shown in the figure, the present invention combines the Unet and Transformer network structures to achieve the recovery of multi-target undersampled signals. During the feature extraction and signal recovery process, this method fully utilizes the 1D-Unet Encoder for dimensionality reduction feature extraction, and the Transformer network is used to capture long-distance dependencies between sequences. Finally, the extracted features are upgraded through the 1D-Unet Decoder to restore the original signal. The Unet-Transformer deep learning network model uses the following loss function to optimize the entire network during training:

[0068] L total =L reconstruction +L regularization ;

[0069] Among them, L reconstruction To restore the signal to the original signal, L regularization is the regularization term, L total is the total loss function;

[0070] The mean square error is used as L reconstruction , the formula is as follows:

[0071]

[0072] Where N is the total number of samples, X i,true is the original signal sample, X i,reconstructed is the signal sample reconstructed by the model, and i is the sample index.

[0073] The 1D-Unet network is a convolutional neural network (CNN) architecture specifically designed for processing one-dimensional signal data. It consists of two main parts: an encoder and a decoder. Through this structure, the network can effectively extract local and global features from the signal and restore the signal. The encoder performs dimensionality reduction on the original input signal through multiple convolutional layers (Conv1D). Each convolutional layer reduces the signal dimension and extracts features through pooling operations. Through multiple layers of convolution-pooling operations, the signal information is gradually compressed into a low-dimensional feature representation. The specific formula is as follows:

[0074] Assume the input signal is Where N is the length of the signal, D is the number of channels of the signal, and after multiple layers of convolution, a low-dimensional feature representation is obtained:

[0075] X encoded =Encoder(X);

[0076] Among them, Encoder(·) is the encoder function, which consists of multiple convolutional layers and pooling layers, X encoded is the low-dimensional output feature representation.

[0077] S4, capture the long-range dependency of the signal through the Transformer network and perform global feature extraction;

[0078] Following the 1D-Unet encoder, the Transformer network is used to capture long-range dependencies in the signal. The Transformer is an architecture capable of processing sequential data and is particularly adept at modeling long-range dependencies between moments in the input signal. In signal recovery tasks, the Transformer can better understand the signal's global information, addressing the limitations of traditional convolutional neural networks in capturing long-range temporal dependencies. Specifically, it uses a multi-head self-attention mechanism to calculate the relationship between each time step in the signal and generate a new feature representation for each time step. The steps are as follows:

[0079] Q,K,V=Linear(X encoded );

[0080] Where Q is the query vector, K is the key vector, V is the value vector, and Linear(·) is the fully connected linear mapping function;

[0081] The attention of each position to all positions is calculated by scaling the dot product attention. The formula is as follows:

[0082]

[0083] Among them, d k is the dimension of query and key, T is the transposition operation, softmax(·) is the normalization function, and Attention(·) is the attention of each position to all positions;

[0084] The new feature representation is obtained through the feedforward neural network FFN. The formula is as follows:

[0085] X transformed =TransformerEncoder(X encoded );

[0086] Among them, X transformed is the encoded feature output, and TransformerEncoder(·) is the Transformer encoder module.

[0087] S5. Use the 1D-Unet decoder to increase the dimension of the low-dimensional features processed by the Transformer network and finally restore the original signal;

[0088] The 1D-Unet Decoder is responsible for increasing the dimension of the low-dimensional features processed by the Transformer network and ultimately restoring the approximate value of the original signal. The solution restores the feature map to the dimension of the original signal through deconvolution layers and upsampling operations, including convolution, deconvolution, and skip connections. The specific formula is as follows:

[0089] X reconstructed =Decoder(X transformed );

[0090] Among them, X reconstructed To restore the original signal, Decoder(·) is the decoder function, which consists of multiple deconvolution layers, upsampling layers and skip connections.

[0091] S6, signal time-frequency conversion, the restored original signal is combined with time-domain reconstruction through time-frequency processing to generate a restored multi-target signal;

[0092] After completing the feature extraction and dimensionality increase of 1D-Unet Encoder, Transformer, and 1D-Unet Decoder, the final recovered signal X reconstructed It will undergo time-frequency processing (such as short-time Fourier transform STFT) combined with time domain reconstruction to generate a restored multi-target signal; the goal of signal recovery is to minimize the information loss caused by undersampling and restore the result closest to the real signal.

[0093] S7. Test the results and conduct performance evaluation. The specific steps are as follows:

[0094] S71. Input the signals in the test set into the trained Unet-Transformer deep learning network model and perform feature extraction, time-frequency analysis, long-distance dependency capture, and signal recovery operations on each signal.

[0095] S72, output the restored signal X reconstructed With the real signal X true Make a comparison;

[0096] S73. Use the evaluation indicators mean square error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM) to quantitatively analyze the quality of the restored signal.

[0097] S74. Display the comparison between the restored signal and the original signal through the time domain waveform and frequency domain spectrum.

[0098] It is worth noting that the contents not elaborated in detail in the present invention are all prior art and are well known to those skilled in the art.

[0099] Therefore, the present invention provides a microwave photon radar multi-signal reconstruction method based on deep learning, which uses deep learning methods to realize data-driven undersampling recovery of multi-target signals, optimizes system performance, overcomes the shortcomings of existing reconstruction methods in detecting and recovering multi-target signals in complex spectral environments, and lays a solid foundation for its wide application in many fields such as national defense, communication networks, imaging and instruments.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A microwave photon radar multi-signal reconstruction method based on deep learning, characterized in that: The following steps are involved: Step S1, generating multi-target microwave radar data through simulation, specifically including multi-target signal generation, multi-channel notching and multi-channel undersampling extraction; Step S2: preprocessing and frequency domain signal extraction of the multi-target microwave radar data, including data noise addition, data standardization and normalization, time-frequency analysis and label design, and dividing the data into a training set and a test set; Step S3: construct a Unet-Transformer deep learning network model and perform feature dimensionality reduction and feature extraction through a 1D-Unet encoder; Step S4: Capture the long-range dependency of the signal through the Transformer network and perform global feature extraction; Step S5: Use the 1D-Unet decoder to increase the dimension of the low-dimensional features processed by the Transformer network and finally restore the original signal; Step S6: signal time-frequency conversion, combining the restored original signal with time-frequency processing and time-domain reconstruction to generate a restored multi-target signal; Step S7: Result testing and performance evaluation.

2. The microwave photon radar multi-signal reconstruction method based on deep learning according to claim 1 is characterized in that: In step S1, the multi-target signal generation includes four types of signals: Chirp, Modsinc, Filtered white noise and LFM; the multi-channel notch introduces different notch strategies for different channels, and the signals in the channels will pass through specific filters or notches to simulate the interference between different targets or signal sources; the multi-channel undersampling extraction dynamically adjusts the number of sampling points according to the frequency characteristics and bandwidth requirements of different signals.

3. The microwave photon radar multi-signal reconstruction method based on deep learning according to claim 1, characterized in that: In step S3, the encoder performs dimensionality reduction processing on the input original signal through multiple convolutional layers Conv1D. Each convolutional layer reduces the dimension of the signal and extracts features through pooling operations. Through multiple layers of convolution-pooling operations, the signal information is gradually compressed into a low-dimensional feature representation. The specific formula is as follows: Assume the input signal is Where N is the length of the signal, D is the number of channels of the signal, and after multiple layers of convolution, a low-dimensional feature representation is obtained: X encoded =Encoder(X); Among them, Encoder(·) is the encoder function, which consists of multiple convolutional layers and pooling layers, X encoded is the low-dimensional output feature representation.

4. The microwave photon radar multi-signal reconstruction method based on deep learning according to claim 1, characterized in that: In step S3, the Unet-Transformer deep learning network model uses the following loss function to optimize the entire network during training: L total =L reconstruction +L regularization ; Among them, L reconstruction To restore the signal to the original signal, L regularization is the regularization term, L total is the total loss function; The mean square error is used as L reconstruction , the formula is as follows: Where N is the total number of samples, X i,true is the real original signal sample, X i,reconstructed is the signal sample reconstructed by the model, and i is the sample index.

5. The microwave photon radar multi-signal reconstruction method based on deep learning according to claim 1, characterized in that: In step S4, the Transformer network is used to capture the long-range dependencies of the signal. Specifically, the multi-head self-attention mechanism Self-Attention is used to calculate the relationship between each time step in the signal and generate a new feature representation for each time step. The steps are as follows: Q,K,V=Linear(X encoded ); Where Q is the query vector, K is the key vector, V is the value vector, and Linear(·) is the fully connected linear mapping function; The attention of each position to all positions is calculated by scaling the dot product attention. The formula is as follows: Among them, d k is the dimension of query and key, T is the transposition operation, softmax(·) is the normalization function, and Attention(·) is the attention of each position to all positions; The new feature representation is obtained through the feedforward neural network FFN. The formula is as follows: X transformed =TransformerEncoder(X encoded ); Among them, X transformed is the encoded feature output, and TransformerEncoder(·) is the Transformer encoder module.

6. The microwave photon radar multi-signal reconstruction method based on deep learning according to claim 1, characterized in that: In step S5, the decoder restores the feature map to the dimension of the original signal through deconvolution layers and upsampling operations, including convolution, deconvolution, and skip connections. The specific formula is as follows: X reconstructed =Decoder(X transformed ); Among them, X reconstructed To restore the original signal, Decoder(·) is the decoder function, which consists of multiple deconvolution layers, upsampling layers and skip connections.

7. The microwave photon radar multi-signal reconstruction method based on deep learning according to claim 1, characterized in that: In step S7, the specific steps of result testing and performance evaluation are as follows: Step S71: Input the signals in the test set into the trained Unet-Transformer deep learning network model, and perform feature extraction, time-frequency analysis, long-range dependency capture, and signal recovery operations on each signal; Step S72: Output the restored signal X reconstructed With the real signal X true Make a comparison; Step S73: Use the evaluation indicators mean square error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM) to quantitatively analyze the quality of the restored signal. Step S74: Display the comparison between the restored signal and the original signal through a time domain waveform diagram and a frequency domain spectrum diagram.

Images