Radar echo signal generation method, electronic equipment and storage medium

Abstract

The invention provides a radar echo signal generation method, electronic equipment and a storage medium, and the method comprises the steps: training an initial generative adversarial network according to a plurality of pre-generated first random noise vectors and a standard radar echo signal, the initial generative adversarial network comprising a generator and a discriminator, the generator is used for generating an initial radar echo signal according to the first random noise vector, and the discriminator is used for comparing the initial radar echo signal with a standard radar echo signal to judge the output effect of the generator; taking the generator at the end of training as a target signal generation network; and inputting the second random noise vector generated in real time into a target signal generation network, and generating a target radar echo signal by the target signal generation network so as to improve the accuracy and diversity of generating the target radar callback signal.

Description

Technical Field

[0001] This application relates to the field of radar technology. Specifically, it relates to a method for generating radar echo signals, an electronic device, and a storage medium. Background Art

[0002] Radar echo signals are key data sources in radar systems and are widely used in fields such as meteorological monitoring and autonomous driving. Digital sample augmentation of radar echo signals can generate additional analog samples to expand the sample library. However, in practical applications, the physical mechanism of radar electromagnetic characteristics is very complex, and there are non-linear time-varying coupling constraints between multiple parameters in radar echo signals.

[0003] Currently, traditional sample augmentation methods for generating harmonics in the time domain are restricted by the physical mechanism constraints of radar electromagnetic characteristics and the non-linear time-varying coupling constraints between parameters, suffering from the drawbacks of low sample fidelity and diversity, and the samples lack interpretability, making it difficult to clearly explain the generation basis of each feature in the samples and the association with the real signal.

[0004] Therefore, there are certain limitations in the digital sample augmentation of radar echo signals in the prior art. Summary of the Invention

[0005] The purpose of this application is to provide a method for generating radar echo signals, an electronic device, and a storage medium to address the practical need for the limitations in the digital sample augmentation of radar echo signals in the prior art.

[0006] To achieve the above objective, the technical solutions adopted in the embodiments of this application are as follows:

[0007] In a first aspect, an embodiment of this application provides a method for generating radar echo signals, the method including:

[0008] Training an initial generative adversarial network according to a plurality of pre-generated first random noise vectors and a standard radar echo signal. The initial generative adversarial network includes a generator and a discriminator. The generator is used to generate an initial radar echo signal according to the first random noise vector, and the discriminator is used to compare the initial radar echo signal with the standard radar echo signal to judge the output effect of the generator;

[0009] Taking the generator at the end of training as a target signal generation network;

[0010] Inputting a real-time generated second random noise vector into the target signal generation network, and generating a target radar echo signal by the target signal generation network.

[0011] As an optional implementation manner, training the initial generative adversarial network according to a plurality of pre-generated first random noise vectors and a standard radar echo signal includes:

[0012] Input the first random noise vector forward into the generator, generate the initial radar echo signal by the generator, output the initial radar echo signal to the discriminator, and determine the loss result of the initial generative adversarial network by the discriminator according to the initial radar echo signal and the standard radar echo signal;

[0013] Adjust the model parameters of the initial generative adversarial network according to the loss result;

[0014] Iteratively execute until the loss result of the initial generative adversarial network meets a preset condition or the number of iterations reaches a preset number, and end the training of the initial generative adversarial network.

[0015] As an optional implementation manner, generating the initial radar echo signal by the generator and outputting the initial radar echo signal to the discriminator includes:

[0016] Input the first random noise vector into the first input layer in the generator;

[0017] Process the first random noise vector sequentially by the first fully connected layer, the second fully connected layer and the first output layer in the generator to obtain the initial radar echo signal, wherein a first activation function is set in the first fully connected layer, and a second activation function is set in the second fully connected layer;

[0018] Output the initial radar echo signal to the discriminator by the first output layer.

[0019] As an optional implementation manner, processing the first random noise vector sequentially by the first fully connected layer, the second fully connected layer and the first output layer in the generator to obtain the initial radar echo signal includes:

[0020] Expand the features of the first random noise vector by the first fully connected layer to obtain first expanded features;

[0021] Correct the first expanded features by the second fully connected layer to obtain first corrected features;

[0022] Normalize the first corrected features by the first output layer to map the first corrected features to a preset interval to obtain the initial radar echo signal.

[0023] As an alternative implementation, the discriminator determines the loss result of the initial generative adversarial network based on the initial radar echo signal and the standard radar echo signal, including:

[0024] Input the initial radar echo signal and the standard radar echo signal into the third fully connected layer in the discriminator, and the third fully connected layer, the fourth fully connected layer, and the second output layer in the discriminator sequentially perform processing to calculate the first loss value of the discriminator and the second loss value of the generator;

[0025] Use the first loss value and the second loss value as the loss result of the initial generative adversarial network.

[0026] As an alternative implementation, the discriminator, based on the initial radar echo signal and the standard radar echo signal, and the third fully connected layer, the fourth fully connected layer, and the second output layer in the discriminator sequentially perform processing to calculate the first loss value of the discriminator and the second loss value of the generator, including:

[0027] The third fully connected layer respectively performs spectral normalization preprocessing on the initial radar echo signal and the standard radar echo signal to obtain the preprocessed initial radar echo signal features and the preprocessed standard radar echo signal features;

[0028] The fourth fully connected layer performs feature discrimination based on the preprocessed initial radar echo signal features and the preprocessed standard radar echo signal features, and extracts the target initial radar echo signal features and the target standard radar echo signal features;

[0029] The second output layer calculates the first loss value of the discriminator and the second loss value of the generator based on the target initial radar echo signal features, the target standard radar echo signal features, and the gradient penalty term.

[0030] As an alternative implementation, adjusting the model parameters of the initial generative adversarial network according to the loss result includes:

[0031] Based on the first loss value of the discriminator, calculate the first gradients corresponding to each layer of the discriminator, and adjust the weights and biases of each layer in the discriminator according to the first gradients;

[0032] Based on the second loss value of the generator, calculate the second gradients corresponding to each layer of the generator, and adjust the weights and biases of each layer in the generator according to the second gradients.

[0033] As an alternative implementation, the method further includes:

[0034] Determine the maximum absolute difference between the first cumulative distribution function corresponding to the standard radar echo signal and the second cumulative distribution function corresponding to the target radar echo signal;

[0035] Evaluate the target radar echo signal according to the maximum absolute difference to obtain an evaluation result;

[0036] Determine whether to continue training the target signal generation network according to the evaluation result.

[0037] In a second aspect, an embodiment of the present application provides a radar echo signal generation device, and the device includes:

[0038] A training module, configured to train an initial generative adversarial network according to a plurality of pre-generated first random noise vectors and a standard radar echo signal. The initial generative adversarial network includes: a generator and a discriminator. The generator is configured to generate an initial radar echo signal according to the first random noise vector, and the discriminator is configured to compare the initial radar echo signal with the standard radar echo signal to judge the output effect of the generator;

[0039] A determination module, configured to use the generator at the end of training as the target signal generation network;

[0040] A generation module, configured to input a second randomly generated noise vector in real time into the target signal generation network, and the target signal generation network generates a target radar echo signal.

[0041] As an optional implementation manner, the training module is specifically configured to:

[0042] Forward the first random noise vector into the generator, and the generator generates the initial radar echo signal, and outputs the initial radar echo signal to the discriminator, and the discriminator determines the loss result of the initial generative adversarial network according to the initial radar echo signal and the standard radar echo signal;

[0043] Adjust the model parameters of the initial generative adversarial network according to the loss result;

[0044] Iteratively execute until the loss result of the initial generative adversarial network meets a preset condition or the number of iterations reaches a preset number, and end the training of the initial generative adversarial network.

[0045] As an optional implementation manner, the training module is specifically configured to:

[0046] Input the first random noise vector into the first input layer of the generator;

[0047] The first fully-connected layer, the second fully-connected layer, and the first output layer in the generator sequentially perform processing to obtain the initial radar echo signal, where a first activation function is set in the first fully-connected layer, and a second activation function is set in the second fully-connected layer;

[0048] The first output layer outputs the initial radar echo signal to the discriminator.

[0049] As an alternative implementation, the training module is specifically configured to:

[0050] The first fully-connected layer performs feature expansion on the first random noise vector to obtain first expanded features;

[0051] The second fully-connected layer corrects the first expanded features to obtain first corrected features;

[0052] The first output layer performs normalization processing on the first corrected features to map the first corrected features to a preset interval, thereby obtaining the initial radar echo signal.

[0053] As an alternative implementation, the training module is specifically configured to:

[0054] The initial radar echo signal and the standard radar echo signal are input into the third fully-connected layer in the discriminator, and the third fully-connected layer, the fourth fully-connected layer, and the second output layer in the discriminator sequentially perform processing to calculate a first loss value of the discriminator and a second loss value of the generator;

[0055] The first loss value and the second loss value are used as the loss result of the initial generative adversarial network.

[0056] As an alternative implementation, the training module is specifically configured to:

[0057] The third fully-connected layer respectively performs spectral normalization preprocessing on the initial radar echo signal and the standard radar echo signal to obtain preprocessed initial radar echo signal features and preprocessed standard radar echo signal features;

[0058] The fourth fully-connected layer performs feature discrimination based on the preprocessed initial radar echo signal features and the preprocessed standard radar echo signal features, and extracts target initial radar echo signal features and target standard radar echo signal features;

[0059] The second output layer calculates a first loss value of the discriminator and a second loss value of the generator according to the target initial radar echo signal characteristics, the target standard radar echo signal characteristics, and the gradient penalty term.

[0060] As an alternative implementation, the training module is specifically configured to:

[0061] Calculate first gradients corresponding to each layer of the discriminator according to the first loss value of the discriminator, and adjust the weights and biases of each layer in the discriminator according to the first gradients;

[0062] Calculate second gradients corresponding to each layer of the generator according to the second loss value of the generator, and adjust the weights and biases of each layer in the generator according to the second gradients.

[0063] As an alternative implementation, the apparatus further includes an evaluation module; the evaluation module is configured to:

[0064] Determine the maximum absolute difference between a first cumulative distribution function corresponding to the standard radar echo signal and a second cumulative distribution function corresponding to the target radar echo signal;

[0065] Evaluate the target radar echo signal according to the maximum absolute difference to obtain an evaluation result;

[0066] Determine whether to continue training the target signal generation network according to the evaluation result.

[0067] In a third aspect, an embodiment of the present application provides an electronic device, including a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device runs, the processor communicates with the memory through the bus, and the processor executes the machine-readable instructions to perform the steps of the radar echo signal generation method described in the first aspect above.

[0068] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, it performs the steps of the radar echo signal generation method described in the first aspect above.

[0069] The beneficial effects of the present application are:

[0070] The present application provides a method for generating radar echo signals, an electronic device, and a storage medium. Based on a first random noise vector and a standard radar echo signal, a generator and a discriminator in an initial generative adversarial network are trained. During the training process, the generator maps the first random noise vector into an initial radar echo signal, and the discriminator determines the authenticity degree of the initial radar echo signal output by the generator based on the comparison result between the standard radar echo signal and the initial radar echo signal output by the generator. Through adversarial training between the generator and the discriminator, the distribution of the initial radar echo signal output by the generator approaches the distribution of the standard radar echo signal. When the training of the initial generative adversarial network ends, the generator is retained as a target signal generation network. The target signal generation network has learned the distribution law of the standard radar echo signal through deep learning to convert a random noise vector into a realistic radar echo signal. The second random noise vector generated in real time is input into the target signal generation network, and the target signal generation network generates a target radar echo signal based on the second random noise vector, such that the distribution law of the target radar echo signal is highly close to that of the standard radar echo signal. By means of deep learning, the temporal correlation features and physical evolution laws of the standard radar echo signal are extracted, the non-linear time-varying coupling constraints between parameters are overcome, the credible augmentation of the radar echo signal is realized, and the fidelity and diversity of the generated target radar echo signal are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present application and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0072] Figure 1 Schematic flowchart of the method for generating radar echo signals provided by the embodiment of the present application Figure 1 ;

[0073] Figure 2 Schematic structural diagram of the initial generative adversarial network provided by the embodiment of the present application;

[0074] Figure 3 Schematic flowchart of the method for generating radar echo signals provided by the embodiment of the present application Figure 2 ;

[0075] Figure 4 Schematic flowchart of the method for generating radar echo signals provided by the embodiment of the present application Figure 3 ;

[0076] Figure 5 Schematic flowchart of the method for generating radar echo signals provided by the embodiment of the present application Figure 4;

[0077] Figure 6 Flow schematic of the radar echo signal generation method provided by the embodiments of the present application Figure 5 ;

[0078] Figure 7 Flow schematic of the radar echo signal generation method provided by the embodiments of the present application Figure 6 ;

[0079] Figure 8 Flow schematic of the radar echo signal generation method provided by the embodiments of the present application Figure 7 ;

[0080] Figure 9 Flow schematic of the radar echo signal generation method provided by the embodiments of the present application Figure 8 ;

[0081] Figure 10 Module structure diagram of the radar echo signal generation device provided by the embodiments of the present application;

[0082] Figure 11 Structural schematic diagram of the electronic device provided by the embodiments of the present application. Detailed implementation manners

[0083] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. It should be understood that the accompanying drawings in the present application are only for the purposes of illustration and description, and are not used to limit the protection scope of the present application. Additionally, it should be understood that the schematic drawings are not drawn to actual scale. The flowcharts used in the present application illustrate the operations implemented according to some embodiments of the present application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical context relationships may be reversed or implemented simultaneously. In addition, those skilled in the art can add one or more other operations to the flowchart or remove one or more operations from the flowchart under the guidance of the content of the present application.

[0084] In addition, the described embodiments are only some of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application usually described and illustrated in the accompanying drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application to be protected, but only represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0085] It should be noted that the term "including" will be used in the embodiments of the present application to indicate the existence of the features stated thereafter, but does not exclude the addition of other features.

[0086] In the field of radar, additional analog samples are generated through the augmentation of digital samples of radar echo signals to expand the sample library. However, in practical applications, the physical mechanism of radar electromagnetic characteristics is very complex, and there are non-linear time-varying coupling constraints among multiple parameters in radar echo signals. Currently, the traditional sample augmentation method for generating harmonics in the time domain is limited by the physical mechanism constraints of radar electromagnetic characteristics and the non-linear time-varying coupling constraints among parameters. The fidelity and diversity of the samples are low, and there is a lack of interpretability, making it difficult to clearly explain the generation basis of each feature in the samples and the association with the real signal. That is to say, there are certain limitations in the digital sample augmentation of radar echo signals in the prior art.

[0087] Based on the above problems, the embodiments of the present application propose a method for generating radar echo signals. The generator and discriminator in the adversarial network are trained based on the first random noise vector and the standard radar echo signal, so that the generator continuously learns the distribution of the standard radar echo signal until the training ends. The target radar echo signal is generated by the trained generator, improving the fidelity and diversity when generating the target radar echo signal.

[0088] Figure 1 Schematic flow of the method for generating radar echo signals provided by the embodiments of the present application Figure 1 The execution subject of this method can be any electronic device with computing and processing capabilities. Such as Figure 1 As shown, this method includes:

[0089] S101. Train the initial adversarial network according to a plurality of pre-generated first random noise vectors and the standard radar echo signal. The initial adversarial network includes: a generator and a discriminator. The generator is used to generate an initial radar echo signal according to the first random noise vector, and the discriminator is used to compare the initial radar echo signal with the standard radar echo signal to judge the output effect of the generator.

[0090] Optionally, obtain the standard radar echo signal and a plurality of first random noise vectors, and input the first random noise vectors and the standard radar echo signal into the initial adversarial network. Among them, the standard radar echo signal can be the measured data in the field or the simulated data obtained through high-precision electromagnetic simulation. The first random noise vector can be a noise vector randomly sampled from a Gaussian distribution in advance, which is a kind of irregular signal. The standard radar echo signal can be used as a reference benchmark when training the initial adversarial network, so that the initial adversarial network learns the distribution of the standard radar echo signal when generating the radar echo signal.

[0091] Among them, in the process of simulating the standard radar echo signal, a dynamic time-domain correlation model of the radar echo signal parameter space is established. Through radio wave propagation delay calculation, fundamental wave generation, echo synthesis, and noise mixing, the process of generating the radar echo signal is simulated, and the standard radar echo signal is obtained and stored.

[0092] Specifically, the transmitted signal (fundamental wave) is generated according to the radar operating frequency band (X-band or S-band). Considering the influence of time delay and noise mixing, physical constraints are established, the radio wave propagation delay and attenuation coefficient are calculated, the radar echo signal is simulated, and the generated time-domain signal waveform is subjected to standardization processing such as amplitude normalization and time-domain alignment to obtain the standard radar echo signal. Among them, the characteristic parameters of the standard radar echo signal include time-domain characteristics (pulse width and amplitude attenuation, etc.), spatial characteristics (distance, azimuth angle, and elevation angle, etc.), noise characteristics (thermal noise and clutter noise, etc.), and scattering characteristics (reflectivity, etc.).

[0093] Figure 2 The structural schematic diagram of the initial generative adversarial network provided by the embodiment of the present application is shown in Figure 2 As shown, the initial generative adversarial network includes a generator and a discriminator. In the training stage, the generator is used to receive the first random noise vector and map the first random noise vector into the time-domain representation (waveform) of the radar echo signal, and output the initial radar echo signal. The discriminator is used to compare the standard radar echo signal with the initial radar echo signal output by the generator, and judge the authenticity degree of the initial radar echo signal output by the generator according to the comparison result.

[0094] During the training process, the training objective of the generator is to generate a more realistic initial radar echo signal, and the training objective of the discriminator is to accurately distinguish the initial radar echo signal from the standard radar echo signal. Through adversarial training between the generator and the discriminator, the distribution of the initial radar echo signal output by the generator is close to the distribution of the standard radar echo signal.

[0095] S102. Use the generator at the end of the training as the target signal generation network.

[0096] Optionally, at the end of the training of the initial generative adversarial network, the model parameters in both the generator and the discriminator have been iteratively updated multiple times. At this time, the generator after multiple updates is separately extracted as the target signal generation network, and the generator after multiple updates already has the ability to generate radar echo signals with high fidelity.

[0097] That is to say, when the initial generative adversarial network training ends, the discriminator is discarded, and only the generator is retained as the target signal generation network. At this time, the generator, i.e., the target signal generation network, extracts the temporal correlation features and physical evolution laws of the standard radar echo signal through deep learning, maintains the physical features of the radar echo signal under the parameter coupling constraint, solves the problem of representing physical laws under high-dimensional coupling in the parameter space, and has learned the internal distribution law of the standard radar echo signal (such as time domain features, spatial features, scattering features, and noise features), and can convert the random noise vector into a realistic radar echo signal.

[0098] S103. Input the second randomly generated noise vector into the target signal generation network, and the target signal generation network generates the target radar echo signal.

[0099] Optionally, in the actual inference stage, input the second randomly generated noise vector into the target signal generation network. The target signal generation network generates the target radar echo signal based on the second randomly generated noise vector. The internal distribution laws such as the time domain features, spatial features, scattering features, and noise features of the target radar echo signal are highly similar to those of the standard radar echo signal, realizing the credible augmentation of the radar echo signal.

[0100] Traditional regularization methods are difficult to be analyzed mathematically in the parameter coupling scenario. The target radar echo signal can be obtained efficiently and accurately through the target signal generation network, thus solving the problem of the failure of the regularization method in the parameter coupling scenario.

[0101] In this embodiment, the generator and discriminator in the initial generative adversarial network are trained based on the first randomly generated noise vector and the standard radar echo signal. During the training process, the generator maps the first randomly generated noise vector into the initial radar echo signal, and the discriminator judges the realism degree of the initial radar echo signal output by the generator based on the comparison result between the standard radar echo signal and the initial radar echo signal output by the generator. Through adversarial training between the generator and the discriminator, the distribution of the initial radar echo signal output by the generator approaches the distribution of the standard radar echo signal. When the initial generative adversarial network training ends, the generator is retained as the target signal generation network. The target signal generation network has learned the distribution law of the standard radar echo signal through deep learning to convert the random noise vector into a realistic radar echo signal. Input the second randomly generated noise vector into the target signal generation network, and the target signal generation network generates the target radar echo signal based on the second randomly generated noise vector, making the distribution law of the target radar echo signal highly similar to that of the standard radar echo signal. Through deep learning, extract the temporal correlation features and physical evolution laws of the standard radar echo signal, overcome the non-linear time-varying coupling constraint between parameters, realize the credible augmentation of the radar echo signal, and improve the fidelity and diversity of the generated target radar echo signal.

[0102] The following will describe in detail the training of the initial generative adversarial network based on a plurality of pre-generated first random noise vectors and a standard radar echo signal.

[0103] Figure 3 The flowchart of the radar echo signal generation method provided by the embodiment of the present application Figure 2 is as Figure 3 shown. In step S101 above, training the initial generative adversarial network based on a plurality of pre-generated first random noise vectors and a standard radar echo signal includes:

[0104] S201. Input the first random noise vector forward into the generator. The generator generates an initial radar echo signal and outputs the initial radar echo signal to the discriminator. The discriminator determines the loss result of the initial generative adversarial network based on the initial radar echo signal and the standard radar echo signal.

[0105] Optionally, continuing to refer to Figure 2 , input the first random noise vector forward into the generator. The generator maps the irregular low-dimensional first random noise vector to a high-dimensional initial radar echo signal with specific rules through a multi-layer neural network and outputs the initial radar echo signal to the discriminator. Exemplarily, the generator generates the initial radar echo signal with the goal of minimizing the probability of being recognized as fake by the discriminator.

[0106] The discriminator receives and compares the standard radar echo signal and the initial radar echo signal output by the generator, determines the probability that the standard radar echo signal is true and the probability that the initial radar echo signal is true, and determines the loss result of the initial generative adversarial network based on the two probability values. Exemplarily, the discriminator outputs 1 (true) for the standard radar echo signal and 0 (false) for the initial radar echo signal by maximizing the correct classification.

[0107] That is to say, during the training process, the discriminator maximizes the discrimination ability between the standard radar echo signal and the initial radar echo signal, and the generator minimizes the discrimination accuracy of the discriminator to form an adversarial game.

[0108] S202. Adjust the model parameters of the initial generative adversarial network according to the loss result.

[0109] Optionally, during the training process, monitor the loss result of the initial generative adversarial network in real time and adjust the model parameters of the initial generative adversarial network according to the loss result, including the parameters of each neural network layer in the generator and the parameters of each neural network layer in the discriminator. The loss result is used to reflect the signal generation effect of the generator and the signal discrimination effect of the discriminator.

[0110] Specifically, the discriminator and the generator are alternately updated according to the loss result. When updating the discriminator, the parameters of each neural network layer in the generator are fixed, and according to the signal discrimination effect of the discriminator reflected by the loss result, the gradient descent algorithm is used to adjust the parameters of each neural network layer in the discriminator. When updating the generator, the parameters of each neural network layer in the discriminator are fixed, and according to the signal generation effect of the generator reflected by the loss result, the gradient descent algorithm is used to adjust the parameters of each neural network layer in the generator.

[0111] S203. Iteratively execute until the loss result of the initial generative adversarial network meets the preset condition or the number of iterations reaches the preset number, and then end the training of the initial generative adversarial network.

[0112] Optionally, repeat the above steps of forward input, initial radar echo signal generation, loss effect determination, and model parameter adjustment to continuously iteratively train the generator and discriminator in the initial generative adversarial network. Each alternating update of the generator and the discriminator can make the signal generation effect of the generator and the signal discrimination effect of the discriminator better, that is, the initial radar echo signal output by the generator is closer to the distribution law of the standard radar echo signal, and the accuracy of the discriminator in distinguishing the initial radar echo signal from the standard radar echo signal is higher.

[0113] Determine whether to stop training the initial generative adversarial network according to whether the loss result of the initial generative adversarial network determined after each iteration meets the preset condition. Specifically, when the loss result of the initial generative adversarial network converges, that is, when the discriminator cannot distinguish the initial radar echo signal from the standard radar echo signal, it means that the signal generation effect of the generator is good. At this time, the model performance of the initial generative adversarial network is saturated, and it is determined to stop training the initial generative adversarial network to maintain the generalization ability and avoid overfitting.

[0114] Alternatively, when the number of iterations of the alternating update of the generator and the discriminator reaches the preset number, stop training the initial generative adversarial network. For example, stop training the initial generative adversarial network after 1000 rounds of alternating update of the generator and the discriminator to avoid wasting resources caused by infinite training, save resources and time, and improve the training efficiency.

[0115] In this embodiment, the first random noise vector is input into the generator in the positive direction. The generator maps the irregular low-dimensional first random noise vector into a high-dimensional initial radar echo signal with specific rules through a multi-layer neural network and outputs it to the discriminator. The discriminator compares the standard radar echo signal and the initial radar echo signal output by the generator to determine the loss result of the initial generative adversarial network. According to the loss result, the model parameters of the initial generative adversarial network are adjusted, including the parameters of each neural network layer in the generator and the parameters of each neural network layer in the discriminator. The generator and the discriminator in the initial generative adversarial network are continuously iteratively trained. When the loss result meets the preset condition or the number of iterations reaches the preset number, the training of the initial generative adversarial network is stopped. Maintain generalization ability, avoid overfitting, save resources and time, and improve training efficiency.

[0116] Hereinafter, the process of generating the initial radar echo signal by the generator and outputting the initial radar echo signal to the discriminator will be described in detail.

[0117] Figure 4 It is a schematic flowchart of the radar echo signal generation method provided by the embodiment of the present application Figure 3 , as Figure 4 shown, in the above step S201, the initial radar echo signal is generated by the generator and output to the discriminator, including:

[0118] S301. Input the first random noise vector into the first input layer of the generator.

[0119] Optionally, continue to refer to Figure 2 , the low-dimensional first random noise vector is input into the first input layer of the generator in the positive direction. The first input layer receives the first random noise vector as the starting point of the generation process. Among them, the distribution of the first random noise vector directly affects the diversity of the initial radar echo signal generated by the generator.

[0120] Exemplarily, the input dimension of the first random noise vector can be 100 dimensions.

[0121] S302. The first fully connected layer, the second fully connected layer and the first output layer in the generator sequentially perform processing to obtain the initial radar echo signal, where a first activation function is set in the first fully connected layer and a second activation function is set in the second fully connected layer.

[0122] Optionally, continue to refer to Figure 2 , the generator further includes a first fully connected layer, a second fully connected layer and a first output layer. The first fully connected layer, the second fully connected layer and the first output layer sequentially perform linear transformation and feature mapping processing on the input first random noise vector to generate a regular high-dimensional initial radar echo signal.

[0123] Among them, a first activation function is set in the first fully connected layer. The first activation function is used to introduce non-linearity after the first fully connected layer performs processing and maintain the gradient flow. A second activation function is set in the second fully connected layer. The second activation function is used to introduce non-linearity after the second fully connected layer performs processing and prevent the gradient from vanishing.

[0124] Through multiple fully connected layers and corresponding activation functions, the generator can learn the distribution characteristics of complex radar echo signals, and can obtain an initial radar echo signal close to the standard radar echo signal without explicit physical modeling.

[0125] S303. Output the initial radar echo signal to the discriminator by the first output layer.

[0126] Optionally, the first output layer outputs the generated initial radar echo signal to the discriminator, so that the discriminator performs signal comparison and discrimination based on the standard radar echo signal and the initial radar echo signal output by the first output layer of the generator.

[0127] In this embodiment, the low-dimensional first random noise vector is positively input to the first input layer in the generator. The first fully connected layer, the second fully connected layer, and the first output layer in the generator sequentially perform linear transformation and feature mapping processing on the input first random noise vector to generate a regular high-dimensional initial radar echo signal. The first activation function set in the first fully connected layer is used to introduce non-linearity after the first fully connected layer performs processing and maintain the gradient flow. The second activation function set in the second fully connected layer is used to introduce non-linearity after the second fully connected layer performs processing and prevent the gradient from vanishing. The first output layer outputs the generated initial radar echo signal to the discriminator, so that the discriminator performs signal comparison and discrimination based on the standard radar echo signal and the initial radar echo signal output by the first output layer of the generator. Through multiple fully connected layers and corresponding activation functions, the generator learns the distribution characteristics of complex radar echo signals and quickly and accurately obtains an initial radar echo signal close to the standard radar echo signal.

[0128] Hereinafter, the process of obtaining the initial radar echo signal by sequentially performing processing by the first fully connected layer, the second fully connected layer, and the first output layer in the generator will be described in detail.

[0129] Figure 5 It is a flowchart of the radar echo signal generation method provided by the embodiment of the present application Figure 4 , as Figure 5 shown, in the above step S302, the process of obtaining the initial radar echo signal by sequentially performing processing by the first fully connected layer, the second fully connected layer, and the first output layer in the generator includes:

[0130] S401. The first fully-connected layer performs feature expansion on the first random noise vector to obtain the first expanded feature.

[0131] Optionally, continue to refer to Figure 2 , the first fully-connected layer expands the low-dimensional first random noise vector into a higher-dimensional feature through a linear transformation, and then introduces non-linearity through the first activation function to obtain the first expanded feature and output it to the second fully-connected layer. That is to say, in the second fully-connected layer, the low-dimensional noise space is mapped to the high-dimensional feature space to generate complex radar echo signals subsequently.

[0132] Among them, the number of neurons in the first fully-connected layer can be 1046 - 4096. The first fully-connected layer performs feature expansion on the first random noise vector, expands the feature expression ability, and extracts the implicit pattern in the first random noise vector. The first activation function can be the first Leaky Rectified Linear Unit (LeakyReLU for short), which avoids the problem of neuron death that is prone to occur in ReLU and maintains the gradient flow.

[0133] S402. The second fully-connected layer corrects the first expanded feature to obtain the first corrected feature.

[0134] Optionally, continue to refer to Figure 2 , the second fully-connected layer performs a second linear transformation on the first expanded feature, and introduces non-linearity through the second activation function to correct the feature representation and refine a more advanced abstract representation, obtaining the first corrected feature and outputting it to the first output layer.

[0135] Among them, the number of neurons in the second fully-connected layer can be the same as or lower than the number of neurons in the first fully-connected layer. For example, the number of neurons in the second fully-connected layer can be 2048. The second activation function can be the second LeakyReLU, which enhances the non-linear expression ability of the third fully-connected layer and prevents the vanishing gradient.

[0136] That is to say, in the third fully-connected layer, the feature is refined, redundant information is suppressed, and the ability to capture key features such as the time-domain feature, spatial feature, scattering feature, and noise feature of the radar echo signal is enhanced.

[0137] S403. The first output layer performs normalization processing on the first corrected feature to map the first corrected feature to a preset interval, obtaining the initial radar echo signal.

[0138] Optionally, the first output layer maps the first corrected feature to the same dimension as the standard radar echo signal (such as the time-domain waveform length) through a linear transformation, and uses a normalization function to constrain the output value within a preset interval, obtaining an initial radar echo signal and outputting it to the discriminator.

[0139] Among them, the normalization function can be the hyperbolic tangent function (Tanh). As a non-linear activation function, the hyperbolic tangent function maps the output value to the preset interval of [-1, 1]. The standard radar echo signal has positive and negative fluctuation characteristics, and the output range of the hyperbolic tangent function is consistent with the positive and negative fluctuation range of the standard radar echo signal, ensuring the physical rationality of the generated initial radar echo signal to match the value range of the standard radar echo signal.

[0140] In this embodiment, the first fully connected layer performs feature expansion on the first random noise vector, obtaining a first expanded feature and outputting it to the second fully connected layer. The second fully connected layer corrects the first expanded feature, obtaining a first corrected feature and outputting it to the first output layer. The first output layer maps the first corrected feature to the same dimension as the standard radar echo signal through a linear transformation, and uses a normalization function to constrain the output value within a preset interval, obtaining an initial radar echo signal and outputting it to the discriminator. The three-layer neural network of the first fully connected layer, the second fully connected layer, and the first output layer in the generator converts the first random noise vector into an initial radar echo signal that conforms to the distribution of the standard radar echo signal through the progressive processing logic of feature expansion, correction, and normalization.

[0141] Next, the process of the discriminator determining the loss result of the initial generative adversarial network based on the initial radar echo signal and the standard radar echo signal will be described in detail.

[0142] Figure 6 It is a flowchart of the radar echo signal generation method provided by the embodiment of the present application Figure 5 , such as Figure 6 shown, in the above step S201, the discriminator determines the loss result of the initial generative adversarial network based on the initial radar echo signal and the standard radar echo signal, including:

[0143] S501. Input the initial radar echo signal and the standard radar echo signal into the third fully connected layer in the discriminator, and the third fully connected layer, the fourth fully connected layer, and the second output layer in the discriminator sequentially perform processing to calculate the first loss value of the discriminator and the second loss value of the generator.

[0144] Optionally, continue to refer to Figure 2, the discriminator includes a third fully-connected layer, a fourth fully-connected layer, and a second output layer. The standard radar echo signal and the initial radar echo signal output by the generator are simultaneously input into the third fully-connected layer of the discriminator. The third fully-connected layer, the fourth fully-connected layer, and the second output layer respectively perform feature extraction and discrimination processing on the standard radar echo signal and the initial radar echo signal in sequence, and calculate the first loss value of the discriminator and the second loss value of the generator.

[0145] Among them, the first loss value of the discriminator is used to evaluate the ability of the discriminator to distinguish the standard radar echo signal from the initial radar echo signal, maximizing the correct classification ability to maximize the correct distinction between the standard radar echo signal and the initial radar echo signal. The second loss value of the generator is used to evaluate the ability of the generator to deceive the discriminator to minimize the discrimination accuracy of the discriminator.

[0146] S502. Take the first loss value and the second loss value as the loss results of the initial generative adversarial network.

[0147] Optionally, the first loss value of the discriminator and the second loss value of the generator together constitute the optimization objective of the initial generative adversarial network. Take the first loss value of the discriminator and the second loss value of the generator as the loss results of the initial generative adversarial network to promote the adversarial evolution of the generator and the discriminator. Decompose the total loss of the initial generative adversarial network into discriminator loss and generator loss, and the optimization strategies of the discriminator and the generator can be adjusted independently to avoid training collapse caused by one party being too strong and improve training stability.

[0148] Specifically, during the adversarial evolution process, the discriminator provides a clear optimization direction for the generator through the output second loss value of the generator, that is, to generate a more realistic initial radar echo signal to reduce the second loss value. The generator forces the discriminator to continuously improve the ability to distinguish the standard radar echo signal from the initial radar echo signal by generating a more realistic initial radar echo signal, forming a virtuous cycle, and finally making the initial radar echo signal generated by the generator approach the feature distribution of the standard radar echo signal.

[0149] In this embodiment, the standard radar echo signal and the initial radar echo signal output by the generator are simultaneously input into the third fully connected layer of the discriminator, and are sequentially processed by the third fully connected layer, the fourth fully connected layer, and the second output layer to calculate the first loss value of the discriminator and the second loss value of the generator. The ability of the discriminator to distinguish between the standard radar echo signal and the initial radar echo signal is evaluated through the first loss value of the discriminator, so as to maximize the correct discrimination between the standard radar echo signal and the initial radar echo signal. The ability of the generator to deceive the discriminator is evaluated through the second loss value of the generator, so as to minimize the discrimination accuracy of the discriminator. The first loss value of the discriminator and the second loss value of the generator are used as the loss results of the initial generative adversarial network. To promote the adversarial evolution of the generator and the discriminator, so that the initial radar echo signal generated by the generator approximates the characteristic distribution of the standard radar echo signal.

[0150] Hereinafter, the process of sequentially processing by the third fully connected layer, the fourth fully connected layer, and the second output layer in the discriminator to calculate the first loss value of the discriminator and the second loss value of the generator will be described in detail.

[0151] Figure 7 Schematic flow of the radar echo signal generation method provided by the embodiment of the present application Figure 6 , as Figure 7 shown, in the above step S501, the third fully connected layer, the fourth fully connected layer, and the second output layer in the discriminator sequentially perform processing to calculate the first loss value of the discriminator and the second loss value of the generator, including:

[0152] S601. The third fully connected layer respectively performs spectral normalization preprocessing on the initial radar echo signal and the standard radar echo signal to obtain the preprocessed initial radar echo signal features and the preprocessed standard radar echo signal features.

[0153] Optionally, for the input initial radar echo signal and standard radar echo signal, spectral normalization is applied in the third fully connected layer to perform spectral norm constraint on the weight matrix of the third fully connected layer. That is, the weight matrix W3 of the third fully connected layer is converted into W3 / σ(W3), where σ(W3) is the spectral norm, that is, the largest singular value of the weight matrix.

[0154] Spectral normalization is introduced in the third fully connected layer of the discriminator, and spectral normalization preprocessing is respectively performed on the initial radar echo signal and the standard radar echo signal to obtain the preprocessed initial radar echo signal features and the preprocessed standard radar echo signal features and output them to the fourth fully connected layer. By constraining the spectral norm (the largest singular value) of the weight matrix of the third fully connected layer in the discriminator, the gradient flow of the third fully connected layer is stabilized, the gradient explosion of the third fully connected layer is suppressed, and the training process is made smoother to facilitate the processing of signals with high dynamic range such as radar echoes.

[0155] S602. The fourth fully connected layer performs feature discrimination based on the preprocessed initial radar echo signal features and the preprocessed standard radar echo signal features, and extracts the target initial radar echo signal features and the target standard radar echo signal features.

[0156] Optionally, the fourth fully connected layer performs feature discrimination on the two types of signal features after spectral normalization preprocessing, extracts high-order features through nonlinear transformation, obtains the target initial radar echo signal features and the target standard radar echo signal features, and outputs them to the second output layer.

[0157] Among them, the target initial radar echo signal features and the target standard radar echo signal features are discriminative features that can distinguish the initial radar echo signal from the standard radar echo signal, such as specific frequency components or time-domain waveform details, etc. By the extracted discriminative features, the sensitivity to the subtle differences between the two types of signals is enhanced.

[0158] Exemplarily, based on the target initial radar echo signal features, potential defects of the initial radar echo signal generated by the capture generator are captured, such as phase discontinuity, Doppler ambiguity, etc. Based on the target standard radar echo signal features, the physical laws of the standard radar echo signal are determined.

[0159] It should be noted that a 25% probability of random inactivation (Dropout) is applied in the fourth fully connected layer, randomly discarding some neurons in the fourth fully connected layer, reducing the co-adaptability between neurons, forcing the discriminator to learn more robust features, breaking the discriminator's dependence on specific neurons, enhancing the generalization ability of the discriminator, and preventing overfitting.

[0160] S603. The second output layer calculates the first loss value of the discriminator and the second loss value of the generator based on the target initial radar echo signal features, the target standard radar echo signal features, and the gradient penalty term.

[0161] Optionally, a gradient penalty term is introduced in the second output layer to perform dynamic gradient penalty on the second output layer, preventing the disappearance of the generator gradient caused by an overly strong discriminator. By the gradient penalty term, the gradient of the discriminator is forced to remain within a preset range, and physical constraints are imposed on the discriminator to ensure that the discriminator does not learn feature differences that violate physical laws. Among them, in the dynamic gradient penalty mechanism, in the initial stage of training, the gradient penalty is small, enabling the discriminator to quickly respond to obvious feature differences so that the generator can quickly learn. In the later stage of training, the gradient penalty is enhanced, restricting the sensitivity of the discriminator to subtle differences and preventing the generator from falling into local optima.

[0162] The second output layer scores the target initial radar echo signal feature and the target standard radar echo signal feature respectively according to the target initial radar echo signal feature and the target standard radar echo signal feature, and calculates the Wasserstein distance between the score of the target initial radar echo signal feature and the score of the target standard radar echo signal feature. According to the dynamic gradient penalty term, the penalty coefficient, and the Wasserstein distance, the first loss value of the discriminator is calculated. According to the score of the target initial radar echo signal feature, the expectation of the discriminator for the initial radar echo signal is determined, and according to the expectation of the discriminator for the initial radar echo signal, the second loss value of the generator is obtained.

[0163] In this embodiment, the third fully connected layer performs spectral normalization preprocessing on the initial radar echo signal and the standard radar echo signal respectively, constrains the spectral norm of the weight matrix of the third fully connected layer, and outputs the preprocessed initial radar echo signal feature and the preprocessed standard radar echo signal feature to the fourth fully connected layer. The fourth fully connected layer discriminates the features of the preprocessed initial radar echo signal feature and the preprocessed standard radar echo signal feature, extracts high-order features through nonlinear transformation, and outputs the target initial radar echo signal feature and the target standard radar echo signal feature to the second output layer. The second output layer introduces a gradient penalty term according to the target initial radar echo signal feature, the target standard radar echo signal feature, and the gradient penalty term, and calculates the first loss value of the discriminator and the second loss value of the generator. The gradient stability during the training of the discriminator is improved, and thus the loss calculation accuracy is improved.

[0164] Next, the process of adjusting the model parameters of the initial generative adversarial network according to the loss result will be described in detail.

[0165] Figure 8 Schematic flow of the radar echo signal generation method provided by the embodiment of the present application Figure 7 , as Figure 8 shown, adjusting the model parameters of the initial generative adversarial network according to the loss result in step S202 above includes:

[0166] S701. Calculate the first gradient corresponding to each layer of the discriminator according to the first loss value of the discriminator, and adjust the weights and biases of each layer in the discriminator according to each first gradient.

[0167] Optionally, continue to refer to Figure 2, backpropagate the first loss value of the discriminator to each neural network layer of the discriminator. Based on the Adaptive Moment Estimation (Adam) optimization algorithm, calculate the first gradients of each neural network layer of the discriminator, and adaptively adjust the learning rates of the weights and biases of each neural network layer in the discriminator according to the first gradients of each neural network layer of the discriminator.

[0168] Fix the weights and biases of each neural network layer in the generator, and adjust the weights and biases of each neural network layer in the discriminator multiple times according to the first gradients of each neural network layer of the discriminator. For example, in one round of alternating updates of the discriminator and the generator, update the discriminator 5 times and then update the generator once, and the learning rates of the weights and biases of each neural network layer in the discriminator are greater than the learning rates of the weights and biases of each neural network layer in the generator to accelerate the evolution of the discriminator's discrimination ability.

[0169] S702. Calculate the second gradients corresponding to each layer of the generator according to the second loss value of the generator, and adjust the weights and biases of each layer in the generator according to each second gradient.

[0170] Optionally, continue to refer to Figure 2 , backpropagate the second loss value of the generator to each neural network layer of the generator. Based on the Adam optimization algorithm, calculate the second gradients of each neural network layer of the generator, and adaptively adjust the learning rates of the weights and biases of each neural network layer in the generator according to the second gradients of each neural network layer of the generator.

[0171] Fix the weights and biases of each neural network layer in the discriminator, and adjust the weights and biases of each neural network layer in the generator once according to the second gradients of each neural network layer of the generator. In one round of alternating updates of the discriminator and the generator, by updating the discriminator multiple times and updating the generator once, ensure the strong discrimination ability of the discriminator, provide reliable gradients for the training evolution of the generator, and the rapid update of the discriminator can better capture the real-time defects of the generator, so that the generator can gradually evolve based on the accurate feedback of the discriminator and reduce ineffective iterations.

[0172] In this embodiment, the first loss value of the discriminator is backpropagated to each neural network layer of the discriminator, and the first gradients of each neural network layer of the discriminator are calculated respectively. The weights and biases of each neural network layer in the generator are fixed, and the weights and biases of each neural network layer in the discriminator are adjusted multiple times according to the first gradients of each neural network layer of the discriminator. The second loss value of the generator is backpropagated to each neural network layer of the generator, and the second gradients of each neural network layer of the generator are calculated respectively. The weights and biases of each neural network layer in the discriminator are fixed, and the weights and biases of each neural network layer in the generator are adjusted once according to the second gradients of each neural network layer of the generator. In one round of alternating update of the discriminator and the generator, by updating the discriminator multiple times and updating the generator once, the strong discrimination ability of the discriminator is ensured, so that the discriminator is updated quickly to better capture the real-time defects of the generator, so that the generator can gradually evolve based on the accurate feedback of the discriminator, reduce ineffective iterations, and improve the training efficiency of the initial generative adversarial network.

[0173] Figure 9 Flow schematic of the radar echo signal generation method provided by the embodiment of the present application Figure 8 , such as Figure 9 shown, the method further includes:

[0174] S801. Determine the maximum absolute difference between the first cumulative distribution function corresponding to the standard radar echo signal and the second cumulative distribution function corresponding to the target radar echo signal.

[0175] Optionally, the standard radar echo signal and the target radar echo signal are sorted according to the latest time sequence respectively, and the Kolmogorov-Smirnov (K-S) test algorithm is used to calculate the maximum absolute difference D between the first cumulative distribution function (CDF) corresponding to the standard radar echo signal and the second CDF corresponding to the target radar echo signal.

[0176] Among them, the smaller the maximum absolute difference D is, the more similar the distributions of the standard radar echo signal and the target radar echo signal are. Exemplarily, if the maximum absolute difference D is 0.05, it means that at any time sequence, the difference between the CDFs of the standard radar echo signal and the target radar echo signal does not exceed 5%.

[0177] S802. Evaluate the target radar echo signal according to the maximum absolute difference to obtain an evaluation result.

[0178] Optionally, according to the number n of target radar echo signals and the preset significance level α of the K-S test, the target critical value D is determined by the look-up table method α,n . According to the target critical value D α,nEvaluate the target radar echo signal according to the magnitude relationship with the maximum absolute difference D to obtain an evaluation result.

[0179] Among them, the significance level α of the K-S test represents the confidence level of the K-S test, and the significance level α of the K-S test is 0.05 (95% confidence level) or 0.01 (99% confidence level).

[0180] S803. Determine whether to continue training the target signal generation network according to the evaluation result.

[0181] Optionally, if the maximum absolute difference D is greater than the target critical value D α,n , it indicates that the difference between the distribution of the standard radar echo signal and the target radar echo signal is relatively large, and it is necessary to continue training the target signal generation network to generate a target radar echo signal that is more similar to the distribution of the standard radar echo signal.

[0182] If the maximum absolute difference D is less than or equal to the target critical value D α,n , it indicates that the difference between the distribution of the standard radar echo signal and the target radar echo signal is relatively small, and there is no need to continue training the target signal generation network.

[0183] In this embodiment, calculate the maximum absolute difference between the first cumulative distribution function corresponding to the standard radar echo signal and the second cumulative distribution function corresponding to the target radar echo signal, and characterize the distribution similarity between the standard radar echo signal and the target radar echo signal through the maximum absolute difference. Determine the target critical value by the look-up table method according to the number of target radar echo signals and the preset test confidence level. Evaluate the target radar echo signal according to the magnitude relationship between the target critical value and the maximum absolute difference. If the maximum absolute difference is greater than the target critical value, it is necessary to continue training the target signal generation network to generate a target radar echo signal that is more similar to the distribution of the standard radar echo signal. If the maximum absolute difference is less than or equal to the target critical value, there is no need to continue training the target signal generation network. Realize the evaluation of the availability rate of the generated target radar echo signal and the test of the signal generation effect of the target signal generation network.

[0184] Based on the same inventive concept, an apparatus for generating a radar echo signal corresponding to the method for generating a radar echo signal is further provided in an embodiment of the present application. Since the principle of solving problems by the apparatus in the embodiment of the present application is similar to the above-mentioned method for generating a radar echo signal in the embodiment of the present application, the implementation of the apparatus can refer to the implementation of the method, and the repeated parts will not be described again.

[0185] Figure 10 For the module structure diagram of the apparatus for generating a radar echo signal provided in an embodiment of the present application, as Figure 10 shown, the apparatus includes:

[0186] A training module 1001 is used to train an initial generative adversarial network according to a plurality of pre-generated first random noise vectors and a standard radar echo signal. The initial generative adversarial network includes: a generator and a discriminator. The generator is used to generate an initial radar echo signal according to the first random noise vector, and the discriminator is used to compare the initial radar echo signal with the standard radar echo signal to judge the output effect of the generator.

[0187] A determination module 1002 is used to use the generator at the end of training as a target signal generation network.

[0188] A generation module 1003 is used to input a second random noise vector generated in real time into the target signal generation network, and the target signal generation network generates a target radar echo signal.

[0189] As an optional implementation manner, the training module 1001 is specifically used for:

[0190] Input the first random noise vector into the generator in the forward direction. The generator generates an initial radar echo signal, outputs the initial radar echo signal to the discriminator, and the discriminator determines the loss result of the initial generative adversarial network according to the initial radar echo signal and the standard radar echo signal.

[0191] Adjust the model parameters of the initial generative adversarial network according to the loss result.

[0192] Iteratively execute until the loss result of the initial generative adversarial network meets a preset condition or the number of iterations reaches a preset number, and then end the training of the initial generative adversarial network.

[0193] As an optional implementation manner, the training module 1001 is specifically used for:

[0194] Input the first random noise vector into the first input layer of the generator.

[0195] The first fully connected layer, the second fully connected layer, and the first output layer in the generator sequentially perform processing to obtain an initial radar echo signal. Among them, a first activation function is set in the first fully connected layer, and a second activation function is set in the second fully connected layer.

[0196] The first output layer outputs the initial radar echo signal to the discriminator.

[0197] As an optional implementation manner, the training module 1001 is specifically used for:

[0198] The first fully connected layer performs feature expansion on the first random noise vector to obtain a first expanded feature.

[0199] The second fully connected layer corrects the first expanded feature to obtain a first corrected feature.

[0200] The first output layer normalizes the first corrected feature to map the first corrected feature to a preset interval, obtaining an initial radar echo signal.

[0201] As an alternative implementation, the training module 1001 is specifically configured to:

[0202] Input the initial radar echo signal and the standard radar echo signal into the third fully connected layer in the discriminator, and the third fully connected layer, the fourth fully connected layer, and the second output layer in the discriminator sequentially perform processing to calculate the first loss value of the discriminator and the second loss value of the generator.

[0203] Use the first loss value and the second loss value as the loss result of the initial generative adversarial network.

[0204] As an alternative implementation, the training module 1001 is specifically configured to:

[0205] The third fully connected layer respectively performs spectral normalization preprocessing on the initial radar echo signal and the standard radar echo signal to obtain preprocessed initial radar echo signal features and preprocessed standard radar echo signal features;

[0206] The fourth fully connected layer performs feature discrimination according to the preprocessed initial radar echo signal features and the preprocessed standard radar echo signal features, and extracts target initial radar echo signal features and target standard radar echo signal features.

[0207] The second output layer calculates the first loss value of the discriminator and the second loss value of the generator according to the target initial radar echo signal features, the target standard radar echo signal features, and the gradient penalty term.

[0208] As an alternative implementation, the training module 1001 is specifically configured to:

[0209] Calculate the first gradient corresponding to each layer of the discriminator according to the first loss value of the discriminator, and adjust the weights and biases of each layer in the discriminator according to each first gradient.

[0210] Calculate the second gradient corresponding to each layer of the generator according to the second loss value of the generator, and adjust the weights and biases of each layer in the generator according to each second gradient.

[0211] As an alternative implementation, the device further includes: an evaluation module 1004. The evaluation module 1004 is used to:

[0212] Determine the maximum absolute difference between the first cumulative distribution function corresponding to the standard radar echo signal and the second cumulative distribution function corresponding to the target radar echo signal.

[0213] Evaluate the target radar echo signal according to the maximum absolute difference to obtain an evaluation result.

[0214] Determine whether to continue training the target signal generation network according to the evaluation result.

[0215] An embodiment of the present application further provides an electronic device, such as Figure 11 shown, which is a schematic structural diagram of the electronic device provided by the embodiment of the present application, including: a processor 111, a memory 112, and a bus 113. The memory 112 stores machine-readable instructions executable by the processor 111 (for example, Figure 10 the execution instructions corresponding to the training module 1001, the determination module 1002, the generation module 1003, and the evaluation module 1004 in the device in

[0216] When the electronic device runs, the processor 111 communicates with the memory 112 through the bus 113. When the machine-readable instructions are executed by the processor 111, the steps of the radar echo signal generation method in the above embodiment are executed.

[0217] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the above-described systems and devices can refer to the corresponding processes in the method embodiments, which will not be elaborated in the present application. In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation. For another example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point, the displayed or discussed couplings or direct couplings or communication connections to each other can be through some communication interfaces. The indirect couplings or communication connections of the devices or modules can be in electrical, mechanical, or other forms.

[0218] In addition, each functional unit in various embodiments of the present application may be integrated into one processing unit, may exist physically alone for each unit, or two or more units may be integrated into one unit. If the function is implemented in the form of a software functional unit and sold or used as an independent product, it may be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art or a part of the technical solution may be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0219] The above are only specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application.

Claims

1. A method for generating radar echo signals, characterized in that, Including: Training an initial generative adversarial network according to a plurality of pre-generated first random noise vectors and a standard radar echo signal, where the initial generative adversarial network includes: a generator and a discriminator, the generator is used to generate an initial radar echo signal according to the first random noise vector, and the discriminator is used to compare the initial radar echo signal with the standard radar echo signal to judge the output effect of the generator; Using the generator at the end of the training as a target signal generation network; Inputting a second random noise vector generated in real time into the target signal generation network, and generating a target radar echo signal by the target signal generation network.

2. The method according to claim 1, wherein The training of the initial generative adversarial network according to the plurality of pre-generated first random noise vectors and the standard radar echo signal includes: Positively inputting the first random noise vector into the generator, generating the initial radar echo signal by the generator, outputting the initial radar echo signal to the discriminator, and determining the loss result of the initial generative adversarial network by the discriminator according to the initial radar echo signal and the standard radar echo signal; Adjusting the model parameters of the initial generative adversarial network according to the loss result; Performing iterative execution until the loss result of the initial generative adversarial network meets a preset condition or the number of iterations reaches a preset number, and ending the training of the initial generative adversarial network.

3. The method according to claim 2, wherein The generating the initial radar echo signal by the generator and outputting the initial radar echo signal to the discriminator includes: Inputting the first random noise vector into the first input layer of the generator; Successively processing by the first fully connected layer, the second fully connected layer and the first output layer in the generator to obtain the initial radar echo signal, where a first activation function is set in the first fully connected layer and a second activation function is set in the second fully connected layer; Outputting the initial radar echo signal to the discriminator by the first output layer.

4. The method according to claim 3, characterized in that, The successively processing by the first fully connected layer, the second fully connected layer and the first output layer in the generator to obtain the initial radar echo signal includes: Performing feature expansion on the first random noise vector by the first fully connected layer to obtain a first expanded feature; Correcting the first expanded feature by the second fully connected layer to obtain a first corrected feature; Performing normalization processing on the first corrected feature by the first output layer to map the first corrected feature to a preset interval to obtain the initial radar echo signal.

5. The method according to claim 2, characterized in that, The determining the loss result of the initial generative adversarial network by the discriminator according to the initial radar echo signal and the standard radar echo signal includes: Inputting the initial radar echo signal and the standard radar echo signal into the third fully connected layer in the discriminator, and successively processing by the third fully connected layer, the fourth fully connected layer and the second output layer in the discriminator to calculate the first loss value of the discriminator and the second loss value of the generator; Use the first loss value and the second loss value as the loss result of the initial generative adversarial network.

6. The method according to claim 5, characterized in that The process of calculating the first loss value of the discriminator and the second loss value of the generator by sequentially processing through the third fully connected layer, the fourth fully connected layer, and the second output layer in the discriminator includes: The third fully connected layer performs spectral normalization preprocessing on the initial radar echo signal and the standard radar echo signal respectively to obtain the preprocessed initial radar echo signal features and the preprocessed standard radar echo signal features; The fourth fully connected layer performs feature discrimination based on the preprocessed initial radar echo signal features and the preprocessed standard radar echo signal features, and extracts the target initial radar echo signal features and the target standard radar echo signal features; The second output layer calculates the first loss value of the discriminator and the second loss value of the generator according to the target initial radar echo signal features, the target standard radar echo signal features, and the gradient penalty term.

7. The method according to claim 5, characterized in that, Adjusting the model parameters of the initial generative adversarial network according to the loss result includes: Calculating the first gradient corresponding to each layer of the discriminator according to the first loss value of the discriminator, and adjusting the weights and biases of each layer in the discriminator according to each of the first gradients; Calculating the second gradient corresponding to each layer of the generator according to the second loss value of the generator, and adjusting the weights and biases of each layer in the generator according to each of the second gradients.

8. The method according to claim 1, wherein The method further includes: Determine the maximum absolute difference between the first cumulative distribution function corresponding to the standard radar echo signal and the second cumulative distribution function corresponding to the target radar echo signal; Evaluate the target radar echo signal according to the maximum absolute difference to obtain an evaluation result; Determine whether to continue training the target signal generation network according to the evaluation result.

9. An electronic device, characterized in that, Including: A processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device runs, the processor communicates with the memory through the bus. The processor executes the machine-readable instructions to perform the steps of the radar echo signal generation method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium. When the computer program is run by the processor, it executes the steps of the radar echo signal generation method according to any one of claims 1 to 8.

Images