Microwave photon radar signal undersampling reconstruction method based on machine learning
Through the undersampling and reconstruction method of microwave photon radar signals based on machine learning, the radar signal is restored using deep learning neural networks, which solves the limitations of existing radar technologies in accuracy and speed, improves system performance and reduces equipment costs.
Patent Information
- Application Number
- CN202510596615.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-09
- Publication Date
- 2025-08-08
AI Technical Summary
The existing radar technology has limitations in precision, signal-to-noise ratio, and calculation speed in compressed sensing signal reconstruction, especially in multi-band reconstructible radars, which are difficult to effectively overcome the limitations of traditional electronic technology.
Using a machine learning-based microwave photon radar signal undersampling reconstruction method, deep learning neural networks, especially Transformer structures, restore the original undersampled signal through signal generation, preprocessing, feature extraction and prediction, and reduce dependence on high-performance devices.
It significantly improves the performance and processing efficiency of the radar system, reduces equipment costs, and reduces dependence on high-performance and high-precision equipment, achieving high-precision signal recovery.
Smart Images
Figure CN120448742A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of information systems, and in particular to a microwave photon radar signal undersampling reconstruction method based on machine learning. Background Art
[0002] Radar, a crucial piece of military equipment in modern warfare, serves as the "eyes" of the navy, army, air force, and space services. It's a crucial means of detecting and tracking threats around the clock, in all weather conditions, and in complex environments. With the advancement of aircraft technology, the characteristics and electromagnetic properties of targets radars need to detect are becoming increasingly complex, necessitating the continuous development and evolution of radar technology and systems to be effective.
[0003] Therefore, realizing new radars that integrate multiple functions and developing and testing multi-band, broadband reconfigurable radars have become important development trends. The rapid development and continued maturity of microwave photonics technology holds great promise for multi-band reconfigurable radars. Compared to traditional electronic technologies, microwave photonics can provide high-frequency, multi-band local oscillators and high-precision, wide-band arbitrary waveform generation. Beamforming based on true optical delay overcomes challenges of beam tilt and aperture crossing in traditional phased arrays. Microwave photonic analog-to-digital conversion can maintain a high effective bit count at high sampling rates.
[0004] Therefore, in order to overcome the limitations of traditional compressed sensing signal reconstruction in terms of accuracy, signal-to-noise ratio and computing speed, the present invention proposes a microwave photon radar signal undersampling reconstruction method based on machine learning. Summary of the Invention
[0005] The purpose of the present invention is to provide a method for undersampling reconstruction of microwave photon radar signals based on machine learning. In view of the limitations of existing compressed sensing signal reconstruction in terms of accuracy, signal-to-noise ratio, and computing speed, this method fully utilizes machine learning methods, especially deep learning methods, to replace traditional compressed sensing technology, realize data-driven undersampling signal recovery, reduce the dependence of optical analog-to-digital conversion on high-performance, high-precision equipment, and reduce equipment costs.
[0006] To achieve the above object, the present invention provides a microwave photon radar signal undersampling reconstruction method based on machine learning, comprising the following steps:
[0007] Step S1: Generate training data by simulating microwave radar data, which specifically includes three parts: signal generation, multi-channel notching, and undersampling extraction;
[0008] Step S2: preprocessing the generated simulated radar data to generate training and test data;
[0009] Step S3: Based on the Transformer structure, a deep learning neural network is constructed to perform feature extraction, mapping, and prediction, obtain the final undersampling mode parameters, and restore the original undersampling signal;
[0010] Step S4: training a deep learning neural network based on the Transformer structure;
[0011] Step S5: Testing the trained Transformer-based deep learning neural network.
[0012] Preferably, in step S1, microwave radar data is simulated to generate training data, and the specific process is as follows:
[0013] Step S11, simulating and generating signal data;
[0014] (1) Setting signal parameters: Setting different pulse width, duty cycle, carrier frequency, and bandwidth parameter descriptions to generate linear frequency modulation signals;
[0015] (2) Setting the signal-to-noise ratio: inserting simulated noise to generate signal data;
[0016] (3) Storing pulse width, duty cycle, carrier frequency, and bandwidth parameters as tag data;
[0017] Step S12, simulating and generating multi-channel notch data;
[0018] Use a multi-channel notch function to perform multi-band filtering on the signal, set the notch center frequency and bandwidth, and store them as label data;
[0019] Step S13, simulating and generating multi-channel under-sampling extraction data;
[0020] The undersampling multiple and the number of undersampling channels are set, and the channel sampling interval is adjusted to achieve undersampling of multi-channel non-uniform radar signals and simulate the multi-channel undersampling signal acquisition of optical analog-to-digital converters.
[0021] Preferably, in step S2, the generated simulated radar data is preprocessed to generate training and test data. The specific process is as follows:
[0022] Step S21, label design for the signal;
[0023] (1) When labeling a signal, the modulation type and modulation parameter labels of the signal are represented as a sequence, arranged in sequence, and identifiers are set to quantize and encode the relationship between each label in the label sequence;
[0024] (2) When quantizing and encoding the signal, for the modulation parameter labels in the label sequence, the continuous values are encoded into discrete values at a certain quantization interval;
[0025] (3) The modulation type label and identifier in the label sequence are directly encoded into the specified discrete value;
[0026] (4) After quantization encoding, a label sequence represented by discrete numerical values is obtained;
[0027] Step S22: Randomly sample the generated simulated radar data, and classify the corresponding missing signals and complete signals into training data and test data.
[0028] Preferably, in step S3, based on the Transformer structure, a deep learning neural network is constructed to perform feature extraction, mapping, and prediction, to obtain the final undersampling mode parameters and restore the original undersampling signal. The specific process is as follows:
[0029] Step S31: In the deep learning neural network, data mapping of the simulated radar signal data is achieved through the embedding layer and the position encoding layer;
[0030] Step S32: The deep feature extraction module uses the Transformer structure and relies on the encoder-decoder framework to input the mapping result into the encoder, which passes through the self-attention layer, feedforward layer, residual connection and normalization layer in sequence to achieve feature extraction;
[0031] The output of the first encoder is passed through the cascaded encoders in sequence. Each encoder has the same structure, including a self-attention layer, a feedforward layer, a residual connection, and a normalization layer, to obtain the query vector and key vector of the last encoder;
[0032] Step S33: Use the query vector and key vector of the last encoder as input to the decoder to obtain the position encoding result;
[0033] Step S34: Input the features extracted and decoded after encoding into two consecutive fully connected layers, obtain the final undersampling mode parameters through full connection calculation, and restore the original undersampling signal.
[0034] Preferably, in step S31, in the deep learning neural network, data mapping of the simulated radar signal data is achieved through the embedding layer and the position encoding layer. The specific process is as follows:
[0035] First, the radar signal sequence is mapped to a low-dimensional feature space through an embedding layer;
[0036] Then, the position encoding layer adds the temporal position information to the embedded radar signal sequence, completes the data mapping of the radar signal, and obtains the mapping result.
[0037] Preferably, in step S32, the mapping result is input into the encoder, and passes through the self-attention layer, the feedforward layer, the residual connection and the normalization layer in the encoder in sequence to realize feature extraction. The specific process is as follows:
[0038] Step S321, self-attention layer;
[0039] A. First, the mapping result is input into the self-attention layer to calculate and generate self-attention features, which are query vector q, key vector k and value vector v respectively;
[0040] B. Calculate the self-attention score by doing the dot product between the query vector of the current time period i and the key vector of other time periods j, which indicates the degree of attention paid to other time periods when encoding the current time period, as shown below:
[0041]
[0042] Among them, q i represents the query vector of the current time period i; k j represents the key vector of other time periods j; s ij represents the self-attention score;
[0043] C. Divide the self-attention score by the square root of the key vector dimension and normalize it through a softmax layer as follows:
[0044]
[0045] Among them, d k represents the key vector dimension; s' ij represents the normalized self-attention score;
[0046] D. Multiply each value vector by the normalized self-attention score and sum them to obtain the final self-attention score, as shown below:
[0047]
[0048] Among them, v i Represents the value vector of the current time period i; z i Represents the final output of the self-attention layer;
[0049] Step S322, feedforward layer;
[0050] The final output of the self-attention layer is fed into the encoder’s feedforward layer, which includes two linear transformation layers and uses the ReLU activation function to obtain the feedforward layer output.
[0051] Step S323: Perform residual connection and normalization operations on the output result of the feedforward layer to obtain the output result of the first encoder.
[0052] Preferably, in step S33, the decoder structure includes two sets of self-attention layers, residual connections and normalization layers, and the specific implementation process includes:
[0053] First, the radar signal feature output extracted by the encoder is input into the first set of self-attention layers, residual connections, and normalization layers of the decoder, and then into the second set of self-attention layers, residual connections, and normalization layers. The query vector and key vector of the second self-attention layer are the query vector and key vector of the last encoder.
[0054] The deep feature extraction module based on the Transformer structure performs position encoding on the input label sequence through the position encoding layer, as shown below:
[0055]
[0056] Among them, PE (1,2a) Indicates the code for even positions; PE (1,2a+1) Represents the encoding of the odd position; a and d represent the index of the dimension in the position encoding and the total dimension of the position encoding respectively; the encoding of the even position and the encoding of the odd position together constitute the result PE of the position encoding.
[0057] Preferably, in step S34, the features extracted and decoded after encoding are input into two consecutive fully connected layers, and the final undersampling mode parameters are obtained through full connection calculation to restore the original undersampling signal. The specific process is as follows:
[0058] Step S341: Calculate the self-attention score of the fully connected layer as follows:
[0059]
[0060] Where Q = (q1, q2, ..., q T ) T represents the query vector of the encoder; K = (k1, k2, ..., k T ) T represents the output key vector of the encoder; V = (v1, v2, ..., v T ) T Represents the output value vector of the encoder; Z=(z1,z2,…,z T ) T Represents the representation vector after self-attention calculation;
[0061] Step S342: The output result of the feedforward layer is as follows:
[0062] FC(z' i )=(z' iW1+b1)W2+b2;
[0063] Among them, z' i Represents the row vector in Z'; W1 and W2 are two trainable parameter matrices; b1 and b2 are trainable bias parameters.
[0064] Preferably, in step S4, a Transformer-based deep learning neural network is trained, and the specific process is as follows:
[0065] The data set obtained in steps S1 and S2 is input into the network, and the radar signal parameter sequence is output through the fully connected layer; the radar signal parameter sequence is obtained by using the real signal parameter label y and the predicted radar signal parameter Construct the mean square error function Loss as follows:
[0066]
[0067] Where N represents the number of samples in the current mini-batch; M i is the number of output parameters of the i-th sample; y ij Represents the true radar signal parameter label of the jth position of the i-th sample; represents the predicted radar signal parameters at the jth position of the i-th sample.
[0068] Preferably, in step S5, the trained Transformer-based deep learning neural network is tested, and the specific process is as follows:
[0069] First, the complex radar signal to be tested is organized into the format of samples in the dataset and input into a deep learning neural network based on the Transformer structure to perform radar signal compressed sensing undersampling recovery.
[0070] Then, the prediction of the label sequence is completed by sequentially reading the files until the recognition results of all undersampled radar signals and the estimation results of the modulation parameters are obtained.
[0071] Therefore, the present invention adopts the above-mentioned microwave photon radar signal undersampling reconstruction method based on machine learning, fully utilizing machine learning methods, especially deep learning methods, to replace traditional compressed sensing technology, realize data-driven undersampling signal recovery, reduce the dependence of optical analog-to-digital conversion on high-performance and high-precision equipment, and reduce equipment costs.
[0072] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0073] Figure 1This is a flow chart of a microwave photon radar signal undersampling reconstruction method based on machine learning in the present invention;
[0074] Figure 2 This is a schematic diagram of the network structure of a microwave photon radar signal undersampling reconstruction method based on machine learning in the present invention. DETAILED DESCRIPTION
[0075] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.
[0076] like Figure 1 As shown, the present invention provides a microwave photon radar signal undersampling reconstruction method based on machine learning, comprising the following steps:
[0077] Step S1: Generate training data by simulating microwave radar data, which specifically includes three parts: signal generation, multi-channel notching, and undersampling extraction;
[0078] Step S2: preprocessing the generated simulated radar data to generate training and test data;
[0079] Step S3: Based on the Transformer structure, a deep learning neural network is constructed to perform feature extraction, mapping, and prediction, obtain the final undersampling mode parameters, and restore the original undersampling signal;
[0080] Step S4: training a deep learning neural network based on the Transformer structure;
[0081] Step S5: Testing the trained Transformer-based deep learning neural network.
[0082] Example
[0083] Step S1: Simulate microwave radar data to generate training data.
[0084] Microwave radar data is simulated to generate training data for subsequent network training. The simulated training data is divided into three parts: signal generation, multi-channel notch filtering, and undersampling extraction. At the same time, the start and end markers of the sequence are set at the beginning and end of the label sequence respectively; separators are set between different labels.
[0085] Step S11: Simulate and generate signal data.
[0086] (1) Setting signal parameters: Taking the linear frequency modulation signal as an example, different pulse width, duty cycle, carrier frequency, and bandwidth parameter descriptions are set to generate the linear frequency modulation signal;
[0087] (2) Setting the signal-to-noise ratio: inserting simulated noise to generate signal data;
[0088] (3) Store the pulse width, duty cycle, carrier frequency, and bandwidth parameters as tag data.
[0089] Step S12: Simulate and generate multi-channel notch data.
[0090] Use a multi-channel notch function to perform multi-band filtering on the signal, set the notch center frequency and bandwidth, and store them as label data.
[0091] Step S13: Simulate and generate multi-channel under-sampling data.
[0092] The undersampling multiple and the number of undersampling channels are set, and the channel sampling interval is adjusted to achieve undersampling of non-uniform radar signals such as multi-channels, thereby simulating the multi-channel undersampling signal acquisition of the optical analog-to-digital converter.
[0093] Step S2: pre-process the generated simulated radar data to generate training and test data.
[0094] Step S21: Design a label for the signal.
[0095] (1) When labeling a signal, the modulation type and modulation parameter labels of the signal are represented in the form of a sequence, arranged in sequence, and an identifier is set to quantize and encode the relationship between each label in the label sequence;
[0096] (2) When quantizing and encoding the signal, for the modulation parameter labels in the label sequence, the continuous values are encoded into discrete values at a certain quantization interval;
[0097] (3) The modulation type label and identifier in the label sequence are directly encoded into the specified discrete value;
[0098] (4) After quantization encoding, a label sequence represented by discrete numerical values is obtained.
[0099] Step S22: Randomly sample the generated simulated radar data, and classify the corresponding missing signals and complete signals into training data or test data.
[0100] Step S3: Based on the Transformer structure, a deep learning neural network is constructed to perform feature extraction, mapping and prediction, and the final undersampling mode parameters are obtained to restore the original undersampling signal, such as Figure 2 shown.
[0101] Step S31: In the deep learning neural network, data mapping of the radar signal is achieved through the embedding layer and the position encoding layer.
[0102] First, the radar signal sequence is mapped to a low-dimensional feature space through an embedding layer;
[0103] Then, the position encoding layer adds the time sequence position information to the embedded radar signal sequence to complete the data mapping of the radar signal and obtain the mapping result S project .
[0104] Step S32: The deep feature extraction module uses the Transformer structure and relies on the encoder-decoder framework to map the result S project The input is passed through the encoder, which passes through the self-attention layer, feedforward layer, residual connection and normalization layer in sequence to achieve feature extraction.
[0105] Step S321, self-attention layer.
[0106] A. First, the mapping result S project Input the self-attention layer and calculate the self-attention features, which are the query vector q, key vector k and value vector v.
[0107] B. Calculate the self-attention score by doing the dot product between the query vector of the current time period i and the key vector of other time periods j, which indicates the degree of attention paid to other time periods when encoding the current time period, as shown below:
[0108]
[0109] Among them, q i represents the query vector of the current time period i; k j represents the key vector of other time periods j; s ij represents the self-attention score.
[0110] C. Divide the self-attention score by the square root of the key vector dimension and normalize it through a softmax layer, as shown below: the resulting structure is all positive and sums to 1;
[0111]
[0112] Among them, d k represents the key vector dimension; s' ij Represents the normalized self-attention score, the results are all positive and the sum is 1.
[0113] D. Multiply each value vector by the normalized self-attention score and sum them to obtain the final self-attention score, as shown below:
[0114]
[0115] Among them, v i Represents the value vector of the current time period i; z i Represents the final output of the self-attention layer.
[0116] Step S322, feedforward layer.
[0117] The final output of the self-attention layer is fed into the feedforward layer of the encoder, which includes two linear transformation layers and uses the ReLU activation function to obtain the output of the feedforward layer.
[0118] Step S323: Perform residual connection and normalization operations on the output result of the feedforward layer to obtain the output result of the first encoder.
[0119] Step S324: Pass the output of the first encoder through the cascaded encoders in sequence. Each encoder has the same structure, including a self-attention layer, a feedforward layer, a residual connection, and a normalization layer, to obtain the query vector and key vector of the last encoder.
[0120] Step S33: Use the query vector and key vector of the last encoder as input to the decoder to obtain the result of position encoding.
[0121] The decoder structure contains two sets of self-attention layers, residual connections and normalization layers.
[0122] First, the radar signal feature output from the encoder is fed into the decoder's first set of self-attention layers, residual connections, and normalization layers, and then into the decoder's second set of self-attention layers, residual connections, and normalization layers. The query vector and key vector of the second self-attention layer are the query vector and key vector of the last encoder.
[0123] The deep feature extraction module based on the Transformer structure performs position encoding on the input label sequence through the position encoding layer, as shown below:
[0124]
[0125]
[0126] Among them, PE (1,2a) Indicates the code for even positions; PE (1,2a+1) represents the encoding of odd positions; a and d represent the index of the dimension in the position encoding and the total dimension of the position encoding respectively. The encoding of even positions and the encoding of odd positions together constitute the result PE of position encoding.
[0127] In step S34, the radar signal is encoded and decoded through a deep feature extraction module based on a Transformer structure, and the features are input into two consecutive fully connected layers. The final undersampling mode parameters are obtained through full connection calculation to restore the original undersampling signal.
[0128] Step S341: Calculate the self-attention score of the fully connected layer as follows:
[0129]
[0130] Where Q = (q1, q2, ..., q T ) T represents the query vector of the encoder; K = (k1, k2, ..., k T ) T represents the output key vector of the encoder; V = (v1, v2, ..., v T ) T Represents the output value vector of the encoder; Z=(z1,z2,…,z T ) T Represents the representation vector after self-attention calculation.
[0131] Step S342: The output result of the feedforward layer is as follows:
[0132] FC(z' i )=(z' i W1+b1)W2+b2;
[0133] Among them, z' i Represents the row vector in Z'; W1 and W2 are two trainable parameter matrices; b1 and b2 are trainable bias parameters.
[0134] Step S4: training a Transformer-based deep learning neural network.
[0135] The data set obtained in steps S1 and S2 is input into the network, and the radar signal parameter sequence is output through the fully connected layer; the radar signal parameter sequence is obtained by using the real signal parameter label y and the predicted radar signal parameter Construct the mean square error function Loss as follows:
[0136]
[0137] Where N represents the number of samples in the current mini-batch; M i is the number of output parameters of the i-th sample; y ij Represents the true radar signal parameter label of the jth position of the i-th sample; represents the predicted radar signal parameters at the jth position of the i-th sample.
[0138] Step S5: Testing the trained Transformer-based deep learning neural network.
[0139] First, the input complex radar signal to be tested is organized into the format of samples in the data set and input into a deep learning neural network based on the Transformer structure to perform compressed sensing undersampling recovery of the radar signal.
[0140] Then, the prediction of the label sequence is completed by sequentially reading the files until the recognition results of all undersampled radar signals and the estimation results of the modulation parameters are obtained.
[0141] Therefore, the present invention adopts the above-mentioned microwave photon radar signal undersampling reconstruction method based on machine learning. In view of the limitations of existing compressed sensing signal reconstruction in terms of accuracy, signal-to-noise ratio and computing speed, it fully utilizes machine learning methods, especially deep learning methods, to replace traditional compressed sensing technology, realize data-driven undersampling signal recovery, reduce the dependence of optical analog-to-digital conversion on high-performance and high-precision equipment, and reduce equipment costs; it not only significantly improves the performance and processing efficiency of the radar system, but also effectively reduces the dependence on high-performance and high-precision equipment while reducing equipment costs.
[0142] 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 signal undersampling reconstruction method based on machine learning, characterized in that: The following steps are involved: Step S1: Generate training data by simulating microwave radar data, which specifically includes three parts: signal generation, multi-channel notching, and undersampling extraction; Step S2: preprocessing the generated simulated radar data to generate training and test data; Step S3: Based on the Transformer structure, a deep learning neural network is constructed to perform feature extraction, mapping, and prediction, obtain the final undersampling mode parameters, and restore the original undersampling signal; Step S4: training a deep learning neural network based on the Transformer structure; Step S5: Testing the trained Transformer-based deep learning neural network.
2. The method for undersampling and reconstruction of microwave photon radar signals based on machine learning according to claim 1, characterized in that: In step S1, microwave radar data is simulated to generate training data. The specific process is as follows: Step S11, simulating and generating signal data; (1) Setting signal parameters: Setting different pulse width, duty cycle, carrier frequency, and bandwidth parameter descriptions to generate linear frequency modulation signals; (2) Setting the signal-to-noise ratio: inserting simulated noise to generate signal data; (3) Storing pulse width, duty cycle, carrier frequency, and bandwidth parameters as tag data; Step S12, simulating and generating multi-channel notch data; Use a multi-channel notch function to perform multi-band filtering on the signal, set the notch center frequency and bandwidth, and store them as label data; Step S13, simulating and generating multi-channel under-sampling extraction data; The undersampling multiple and the number of undersampling channels are set, and the channel sampling interval is adjusted to achieve undersampling of multi-channel non-uniform radar signals and simulate the multi-channel undersampling signal acquisition of optical analog-to-digital converters.
3. The method for undersampling and reconstruction of microwave photon radar signals based on machine learning according to claim 1, characterized in that: In step S2, the generated simulated radar data is preprocessed to generate training and test data. The specific process is as follows: Step S21, label design for the signal; (1) When labeling a signal, the modulation type and modulation parameter labels of the signal are represented in the form of a sequence, arranged in sequence, and an identifier is set to quantize and encode the relationship between each label in the label sequence; (2) When quantizing and encoding the signal, for the modulation parameter labels in the label sequence, the continuous values are encoded into discrete values at a certain quantization interval; (3) The modulation type label and identifier in the label sequence are directly encoded into the specified discrete value; (4) After quantization encoding, a label sequence represented by discrete numerical values is obtained; Step S22: Randomly sample the generated simulated radar data, and classify the corresponding missing signals and complete signals into training data and test data.
4. The method for undersampling and reconstruction of microwave photon radar signals based on machine learning according to claim 1, characterized in that: In step S3, based on the Transformer structure, a deep learning neural network is constructed to perform feature extraction, mapping, and prediction to obtain the final undersampling mode parameters and restore the original undersampling signal. The specific process is as follows: Step S31: In the deep learning neural network, data mapping of the simulated radar signal data is achieved through the embedding layer and the position encoding layer; Step S32: The deep feature extraction module uses the Transformer structure and relies on the encoder-decoder framework to input the mapping results into the encoder. In the encoder, the results pass through the self-attention layer, feedforward layer, residual connection and normalization layer in sequence to achieve feature extraction. The output of the first encoder is passed through the cascaded encoders in sequence. Each encoder has the same structure, including a self-attention layer, a feedforward layer, a residual connection, and a normalization layer, to obtain the query vector and key vector of the last encoder; Step S33: Use the query vector and key vector of the last encoder as input to the decoder to obtain the position encoding result; Step S34: Input the features extracted and decoded after encoding into two consecutive fully connected layers, obtain the final undersampling mode parameters through full connection calculation, and restore the original undersampling signal.
5. The method for undersampling and reconstruction of microwave photon radar signals based on machine learning according to claim 4, characterized in that: In step S31, in the deep learning neural network, data mapping of the simulated radar signal data is achieved through the embedding layer and the position encoding layer. The specific process is as follows: First, the radar signal sequence is mapped to a low-dimensional feature space through an embedding layer; Then, the position encoding layer adds the temporal position information to the embedded radar signal sequence, completes the data mapping of the radar signal, and obtains the mapping result.
6. The method for undersampling and reconstruction of microwave photon radar signals based on machine learning according to claim 4, characterized in that: In step S32, the mapping result is input into the encoder, and passes through the self-attention layer, feedforward layer, residual connection and normalization layer in the encoder to realize feature extraction. The specific process is as follows: Step S321, self-attention layer; A. First, the mapping result is input into the self-attention layer to calculate and generate self-attention features, which are query vector q, key vector k and value vector v respectively; B. Calculate the self-attention score by doing the dot product between the query vector of the current time period i and the key vector of other time periods j, which indicates the degree of attention paid to other time periods when encoding the current time period, as shown below: Among them, q i represents the query vector of the current time period i; k j represents the key vector of other time periods j; s ij represents the self-attention score; C. Divide the self-attention score by the square root of the key vector dimension and normalize it through a softmax layer as follows: Among them, d k represents the key vector dimension; s' ij represents the normalized self-attention score; D. Multiply each value vector by the normalized self-attention score and sum them to obtain the final self-attention score, as shown below: Among them, v i Represents the value vector of the current time period i; z i Represents the final output of the self-attention layer; Step S322, feedforward layer; The final output of the self-attention layer is fed into the feedforward layer of the encoder, which includes two linear transformation layers and uses the ReLU activation function to obtain the output of the feedforward layer; Step S323: Perform residual connection and normalization operations on the output result of the feedforward layer to obtain the output result of the first encoder.
7. The method for undersampling and reconstruction of microwave photon radar signals based on machine learning according to claim 4, characterized in that: In step S33, the decoder structure includes two sets of self-attention layers, residual connections and normalization layers. The specific implementation process includes: First, the radar signal feature output extracted by the encoder is input into the first set of self-attention layers, residual connections, and normalization layers of the decoder, and then into the second set of self-attention layers, residual connections, and normalization layers. The query vector and key vector of the second self-attention layer are the query vector and key vector of the last encoder. The deep feature extraction module based on the Transformer structure performs position encoding on the input label sequence through the position encoding layer, as shown below: Among them, PE (1,2a) Indicates the code for even positions; PE (1,2a+1) Represents the encoding of the odd position; a and d represent the index of the dimension in the position encoding and the total dimension of the position encoding respectively; the encoding of the even position and the encoding of the odd position together constitute the result PE of the position encoding.
8. The method for undersampling and reconstruction of microwave photon radar signals based on machine learning according to claim 4, characterized in that: In step S34, the features extracted and decoded are input into two consecutive fully connected layers, and the final undersampling mode parameters are obtained through full connection calculation to restore the original undersampling signal. The specific process is as follows: Step S341: Calculate the self-attention score of the fully connected layer as follows: Where Q = (q1, q2, ..., q T ) T represents the query vector of the encoder; K = (k1, k2, ..., k T ) T represents the output key vector of the encoder; V = (v1, v2, ..., v T ) T Represents the output value vector of the encoder; Z=(z1,z2,…,z T ) T Represents the representation vector after self-attention calculation; Step S342: The output result of the feedforward layer is as follows: FC(z' i )=(z' i W1+b1)W2+b2; Among them, z' i Represents the row vector in Z'; W1 and W2 are two trainable parameter matrices; b1 and b2 are trainable bias parameters.
9. The method for undersampling and reconstruction of microwave photon radar signals based on machine learning according to claim 1, characterized in that: In step S4, the Transformer-based deep learning neural network is trained. The specific process is as follows: The data set obtained in steps S1 and S2 is input into the network, and the radar signal parameter sequence is output through the fully connected layer; the radar signal parameter sequence is obtained by using the real signal parameter label y and the predicted radar signal parameter Construct the mean square error function Loss as follows: Among them, N represents the number of samples in the current mini-batch; M i is the number of output parameters of the i-th sample; y ij Represents the true radar signal parameter label of the jth position of the i-th sample; represents the predicted radar signal parameters at the jth position of the i-th sample.
10. The method for undersampling and reconstruction of microwave photon radar signals based on machine learning according to claim 1, characterized in that: In step S5, the trained Transformer-based deep learning neural network is tested. The specific process is as follows: First, the complex radar signal to be tested is organized into the format of samples in the dataset and input into a deep learning neural network based on the Transformer structure to perform radar signal compressed sensing undersampling recovery. Then, the prediction of the label sequence is completed by sequentially reading the files until the recognition results of all undersampled radar signals and the estimation results of the modulation parameters are obtained.