Radar single target detection method, system, and device based on convolutional neural network

Abstract

The application discloses a radar single-target detection method, system and device based on a convolutional neural network, and belongs to the technical field of radar target detection. The radar single-target detection method first carries out preprocessing on radar echo data, including orthogonal phase detection pulse pressure, to obtain a one-dimensional fast time distance sequence of a target. Then, the one-dimensional fast time distance sequence is appropriately deformed as an input data image, and a full zero vector and a zero-one vector are used as output labels for classification training. The convolutional neural network is used for training, and the hyperparameters are appropriately adjusted, so that the error between the input and the output labels meets the requirements, and the training network parameters are obtained. Finally, the network parameters are used for target detection and distance parameter estimation on received data. Through simulation experiments, compared with a traditional radar target detection method, the method has obvious advantages in accuracy and real-time performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar target detection, and particularly relates to a radar single-target detection method, system and device based on a convolutional neural network. BACKGROUND

[0002] A neural network is a highly complex nonlinear learning system that can be used for predictive modeling, adaptive control and training through a data set. Artificial neural networks have been successfully applied to autonomous robots, speech recognition, image analysis and adaptive control in the fields of radar and communication. Common neural networks include BP neural networks, radial basis neural networks, perceptron neural networks, self-organizing neural networks and feedback neural networks. A deep convolutional neural network (DCNN) is a special neural network composed of multiple convolutional layers, pooling layers and fully connected layers. This network structure performs well in processing high-dimensional data such as images and videos, because it has the ability to automatically learn image features. The LeNet-5 model is the pioneer of deep convolutional neural networks, and it is widely used in tasks such as handwritten digit recognition. The AlexNet model is a landmark work in the field of deep convolutional neural networks, leading a new round of development in the field of deep learning. The design ideas and network structures of the VGG and GoogleNet / Inception, YOLO, ResNet models are different, but they have all made important contributions to the development of convolutional neural networks.

[0003] Radar target detection is one of the core functions of a radar system, and its task is to extract the position, speed, shape and other features of a target by processing the received signals, so as to realize the identification and tracking of the target. In order to solve these problems, researchers have proposed many radar target detection algorithms, such as constant false alarm rate detection algorithms, target detection in complex environments, sea surface target detection algorithms, passive multi-base radar target detection methods and cooperative target detection methods based on FDA arrays. In addition, with the development of artificial intelligence and machine learning technologies, researchers have also begun to apply deep learning and other neural network technologies to radar target detection, and have achieved good results.

[0004] Sun Ning et al. from the signal prediction and feature classification two angles, respectively, using long short-term memory network (LSTM) and convolutional neural network (CNN) for signal time series amplitude information processing, for offshore target one-dimensional sequence radar signal detection, this method does not need to assume the background statistical characteristics in advance, and has stronger generalization ability. Liu Junwei et al. proposed a deep learning radar target detection method based on conditional generative adversarial network, which achieved good detection effect. Chen Xiaolong et al. designed a double-channel convolutional neural network detector to jointly extract amplitude features and time-frequency features from the amplitude sequence and time-frequency spectrum of the return unit to improve the performance of the detector, and the false alarm is controllable. Wang Jingang et al. designed an offshore target detector based on attention mechanism to extract high-dimensional difference features of targets and clutter from the echo to realize detection. Chen Simin converts the phase of the return unit into a graph and uses a graph convolutional neural network to realize target detection. Wang Zhi fei et al. use the data features between echo frames and frames to realize radar target detection by using deep learning. In order to utilize the sequence feature difference between target and clutter to enhance the detection performance, Wan Hao proposed a detection method based on LSTM network. Wang Yumiao uses a complex UNet network to realize a sea surface small target detection method with controllable false alarm rate.

[0005] Although the above methods can realize radar single target detection, due to the complexity of the network structure, the inference time of the deep learning model may be long, and the real-time performance may be insufficient when performing radar target detection, and the real-time performance of the radar system is extremely high, which will inevitably affect the response speed and real-time decision-making ability of the radar system; in addition, the detection and estimation in the above method are separated, and the radar target detection and estimation cannot be realized synchronously, which will cause the total time required for target detection and estimation to be long. SUMMARY

[0006] In view of the above problems, the present application aims to provide a radar single target detection method, system and device based on convolutional neural network, which can realize the detection of radar single target and the estimation of target distance by using multi-layer two-dimensional convolutional neural network after orthogonal phase detection processing of radar echo data, and has obvious advantages in accuracy and real-time performance.

[0007] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0008] The radar single target detection method based on convolutional neural network comprises the following steps:

[0009] S1: pre-processing the radar echo data to obtain one-dimensional fast time distance sequence of the target;

[0010] S2: Construct a convolutional neural network for detecting radar single targets and train it;

[0011] S3: Use the network parameters trained in step S2 to perform target detection and distance parameter estimation on the preprocessed data in step S1.

[0012] Further, the specific operation of step S1 includes the following steps:

[0013] S101: Establish a return signal model after the radar transmitted signal encounters a target.

[0014] S102: Perform down-conversion processing on the return signal of the radar, and then perform quadrature phase detection to obtain the echo signal model, i.e., the one-dimensional fast time distance sequence of the target.

[0015] Further, the echo signal model in step S102 is represented as

[0016] x'(t) = x(t) + n(t)

[0017] where x'(t) is the echo signal, and n(t) is the noise; A is the amplitude of the echo; u(t-t r ) is the delay of the echo signal; is the time delay caused by the target, R t is the distance of the target, and c is the speed of light; is the Doppler frequency of the target, v r is the radial velocity of the target, and λ is the wavelength of the radar.

[0018] Further, the convolutional neural network in step S2 includes six convolutional layers and two fully connected layers, each convolutional layer has 256 convolutional kernels, the first two convolutional layers have a kernel size of 3x3, and the last four convolutional layers have a kernel size of 2x2. After each convolutional layer, there is a batch normalization layer and a ReLU layer. After all the convolutional layers, there are two fully connected layers. After the first fully connected layer, there is a ReLU activation function and a dropput layer. After the second fully connected layer, a sigmoid activation function is used to induce classification probability and output.

[0019] Further, the convolutional layer network parameters of the convolutional neural network are where W e represents a convolutional kernel with dimensions c e x f e x f e x n e , c e represents the number of input channels, f e represents the length and width of the convolutional kernel, and n e represents the number of convolutional kernels, i.e., the number of output channels, and be denotes a bias vector with dimension n e .

[0020] Further, the network parameters of the full connection layer of the convolutional neural network are wherein W 0f denotes a weight matrix with dimension c f ×n f , c f denotes the number of input channels, n f denotes the number of full connection layer neurons, that is, the number of output channels, and b 0f denotes a bias vector with dimension n f .

[0021] Further, the input data and the output label pair of the convolutional neural network are represented as

[0022] D train ={(k 1 , Y 1 ), (k 2 , Y 2 ), …, (k l , Y l )}

[0023] wherein Y denotes the network input of the training sample capacity, k is a label, and l is the number of samples, if a target exists, k is a zero-one vector, and the position of the non-zero element is the target distance, and if no target is detected, k is a zero vector.

[0024] Further, the present application also includes a radar single target detection system based on a convolutional neural network, the detection system comprising a radar data preprocessing module and a convolutional neural network module, the radar data preprocessing module being used for preprocessing radar echo data to obtain a one-dimensional fast time distance sequence of a target, and the convolutional neural network module being used for detecting a radar single target.

[0025] The radar data preprocessing module and the convolutional neural network module are implemented based on the detection method as described above.

[0026] Further, the present application also includes an electronic device, comprising at least one processor, and a memory connected in communication with the processor, wherein the memory stores instructions executable by the processor, and the instructions are executed by the processor to enable the processor to execute the detection method as described above.

[0027] The present application has the following beneficial effects:

[0028] 1. The radar single target detection method based on a convolutional neural network is proposed in the paper. First, the radar echo data is preprocessed, including orthogonal phase detection pulse compression, etc. to obtain the one-dimensional fast time distance sequence of the target. Then, the input data image is appropriately deformed, and the all-zero vector and the zero-one vector are used as output labels for classification training. The convolutional neural network is trained, and the hyperparameters are appropriately adjusted to make the input and output label error meet the requirements, and the training network parameters are obtained. Finally, the network parameters are used for target detection and distance parameter estimation of the received data. The radar target detection method in the present application is experimentally verified, and the results show that the method has high performance in radar target detection, can automatically learn the features in the image, avoids the tediousness and subjectivity of manual feature extraction. At the same time, the method can also improve the accuracy and recall rate of radar target detection, and the real-time performance. Therefore, the method has wide application prospect, and can provide more accurate and faster target detection capability for radar systems.

[0029] 2. The radar single target detection method in the present application can realize the detection and estimation of radar targets at the same time, greatly saving the total time of target detection and estimation, and improving the accuracy accordingly. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 It is a radar signal echo image of the single target fast time dimension in the present application.

[0031] Figure 2 It is a convolutional neural network structure diagram in the present application.

[0032] Figure 3 It is a learning rate change diagram during the training of the convolutional neural network in the simulation experiment of the present application.

[0033] Figure 4 It is a loss function change diagram during the training of the convolutional neural network in the simulation experiment of the present application.

[0034] Figure 5 It is the no-target induced probability output result in the simulation experiment of the present application.

[0035] Figure 6 It is the target induced probability output result in the simulation experiment of the present application.

[0036] Figure 7 It is the 50-time target distance estimation result in the simulation experiment of the present application. DETAILED DESCRIPTION

[0037] In order to make those skilled in the art better understand the technical scheme of the present application, the technical scheme of the present application is further described below in combination with the drawings and examples.

[0038] Example one:

[0039] Embodiment one provides a radar single target detection method based on a convolutional neural network, the detection method comprising the following steps:

[0040] S1: pre-processing radar echo data to obtain a one-dimensional fast time-distance sequence of a target;

[0041] The radar uses a transmitter to emit a radar signal, and in the process of signal propagation from the radar to the target and return to the radar, the signal is affected by factors such as atmospheric refraction, attenuation, multipath effect, etc., resulting in changes in the amplitude and phase of the signal. After the radar receiver receives the echo signal, it needs to be amplified, filtered, mixed, etc. to extract the distance, speed, etc. of the target. Specifically, the pre-processing of the radar echo data in the present application comprises the following steps:

[0042] S101: establishing a return signal model after the radar transmitted signal encounters a target;

[0043] Let the radar transmitted signal e(t) be

[0044]

[0045] Where u(t) is the baseband transmitted signal, which can be a linear frequency modulation signal, a phase encoding signal, and the most commonly used linear frequency modulation signal pulse train; f0 is the carrier frequency; is the initial phase. Then the return signal r(t) of the radar after the transmitted signal encounters a target is:

[0046]

[0047] Where A is the amplitude of the echo; u(t-t r ) is the delay of the echo signal; is the time delay caused by the target, which is the parameter to be solved, R t is the target distance, c is the speed of light; is the Doppler frequency of the target, which is also the parameter of interest of the radar system, v r is the radial velocity of the target, and λ is the radar operating wavelength.

[0048] S102: down-conversion processing of the return signal of the radar, and then quadrature phase detection to obtain an echo signal model, i.e. a one-dimensional fast time-distance sequence of a target;

[0049] After obtaining the return signal of the radar, the radar system needs to be down-converted by one or more stages, and then through quadrature phase detection, the echo signal x(t) is obtained

[0050]

[0051] If there is noise, the echo signal is further represented as

[0052] x'(t) = x(t) + n(t) (4)

[0053] It should be noted that if x'(t) is a compressible pulse train, the traditional range-Doppler map can be obtained by pulse compression, matched filter set, and modulus value calculation. In order to simplify, the present application only considers the result after one-dimensional fast-time pulse compression. In order to facilitate processing, it is assumed that the target occupies only one range cell, and it is assumed that the noise is subject to Gaussian white noise. Assuming that the range unit is 100, and the target position is in the 50th range unit, the fast-time dimension signal of the received sample is as shown in the accompanying Figure 1 The target of the present application is to judge whether there is a target, and if there is a target, to estimate the range unit of the target.

[0054] Step S2: constructing a convolutional neural network for detecting a radar single target, and training;

[0055] The input layer of the convolutional neural network is used to obtain the input data of the convolutional neural network, and the characteristics of itself are retained. In the convolutional neural network for processing images, the input layer converts image data into matrix pixel data input into the neural network, and the first two parameters in the matrix represent the length and width of the input image, and the third parameter is the depth representing the number of channels of the image input. The role of the input layer is to perform some preliminary processing operations on the input image, including normalization, mean removal, etc. The output layer is the last layer of the convolutional neural network, which is used to output the prediction result of the model. In general, the input and output of the convolutional neural network are for the training and prediction tasks of the model.

[0056] In the present application, a supervised learning method (network training based on pre-given data and labels) is adopted, and according to the classical one-dimensional fast-time dimension echo data of formula (4), the data vector r is obtained by sampling. Assuming that the detection signal is M range units, the input data dimension is Mx1x1, that is, the lengthxwidth is Mx1, and the number of channels is 1. It should be noted that the present application adopts a convolutional neural network, and the length and width of the data are preferably in the form of a matrix, therefore, in the present application, the data is changed into an NxNx1 square data Y (in fact, this data can be a range-Doppler two-dimensional image) as a multi-channel input of the network. The output is an Mx1 signal vector, and if there is no target, the output probability is relatively small, and here we define that if it is not more than 0.5, there is no target, and if it is more than 0.5, the position of the maximum element is the distance value of the target. The target distance traverses all the range units, and is randomly extracted according to the ratio of 8:2 to make a training set and a test set, and the number of samples is l.

[0057] In training, the position of the non-zero element in the signal vector (target distance) is taken as the label, denoted as k; for example, if the received signal sampling data of the radar fast time dimension is 10 distance units, and there is a target at the 7th distance unit, then k = [0, 0, 0, 0, 0, 0, 1, 0, 0], and if there is no target, the label k = [0, 0, 0, 0, 0, 0, 0, 0, 0].

[0058] The input data and the label pair of the training set of the network are represented as follows:

[0059] D train = {(k 1 ,Y 1 ), (k 2 ,Y 2 ), …, (k l ,Y l )} (5)

[0060] Where Y represents the network input of the training sample capacity, k is the label, and l is the sample number.

[0061] The construction of the target detection network usually includes the following steps: determining the network structure: usually using the network of transfer learning, based on some existing classic network structure, such as AlexNet, VGG, ResNet, etc., adding a detection head at the end of the network. Loss function design: the commonly used loss functions for target detection tasks include cross-entropy loss and L1 loss, etc. Training and optimization: use the training data to train the target detection network, and continuously optimize the parameters of the network through the back propagation algorithm. In the training process, some optimization algorithms such as gradient descent, Adam, etc. can be used to accelerate the convergence of the network. Test and evaluation: after training is completed, the target detection network needs to be tested and evaluated.

[0062] The convolutional neural network structure for detecting radar single targets in the application is shown in the accompanying Figure 2 The CNN with multiple layers of two-dimensional convolution structure is used for single target detection and target distance estimation. The network has six convolution layers and two fully connected layers. Each convolution layer has 256 convolution kernels, the first two convolution layers have a kernel size of 3x3, and the last four have a kernel size of 2x2. Each convolution layer is followed by a batch normalization layer and a ReLU layer; after all the convolution layers, there are two fully connected layers, followed by a ReLU activation function and a dropput layer after the first fully connected layer, and a sigmoid activation function is used to induce classification probability and output after the second fully connected layer. Since each convolution layer uses edge zero padding operation, only the channel number is changed, and the length and width of the data remain unchanged.

[0063] The network parameters of the convolution layer in the application are where W e represents the dimension ce ×f e ×f e ×n e the convolution kernel, c e represents the number of input channels, f e represents the length and width of the convolution kernel, n e represents the number of convolution kernels (i.e., the number of output channels), b e represents a bias vector with a dimension of n e . The convolution operation is denoted by , and the operation of each convolution layer is as follows:

[0064] ① The convolution layer Conv-1 is used to extract features from the input data, and a ReLU activation function is used, where c e = 3, n e = 256. The specific operation can be represented as:

[0065]

[0066] where Y represents the network input of the training sample capacity.

[0067] ② The convolution layers Conv-2 to Conv-5 implement nonlinear mapping of features, and a ReLU activation function is used, where c e = 256, n e = 256. The specific operation can be represented as:

[0068]

[0069] ③ The convolution layer Conv-6 is a data reconstruction layer, and a data output with a dimension of M x M x 256 is obtained, where c e = 256, n e = 256. The specific operation can be represented as:

[0070]

[0071] The network parameters of the fully connected layer are as follows: where W 0f represents a weight matrix with a dimension of c f x n f , c f represents the number of input channels, n f represents the number of neurons of the fully connected layer (i.e., the number of output channels), b 0f represents a bias vector with a dimension of n f . The matrix multiplication operation is denoted by

[0072] ① The fully connected layer FC-1 is used to integrate convolution layer information, and a ReLU activation function is used, where c f= M x M x 256, n f = 1024. The specific operation can be represented as:

[0073]

[0074] The full connection layer FC-2 is used to induce classification probability and output, and adopts a sigmoid activation function, at this time c f = 1024, n f = N. The specific operation can be represented as:

[0075]

[0076] In the formula, is the actual output of the network obtained by the training data Y through the DL network, The value of has two cases, one is the label without target, and the other is the label with target. In the network training process, the cross entropy of the label and the network output is used as the loss function, that is:

[0077]

[0078] In addition, due to the large number of DL network parameters and the complex optimization problem, during the network training process, the learning rate should be gradually reduced with the increase of the number of iterations, so that the network parameters gradually converge to the optimal or near-optimal.

[0079] In order to prevent the problem of network overfitting, the Dropout strategy is adopted in the application, that is, to randomly remove neurons with a certain probability, so as to alleviate the dependence of the network on the distribution of training set data. Through training, the actual output of the network is more and more close to k, when the network training is completed, the test is carried out, and the test sample is input to the DL network in the form of multi-channel data Y, the network output is obtained and whether it is close to the real label k is verified, so as to judge the network training level.

[0080] In addition, the back propagation algorithm is used in the application to minimize the loss function to train the network parameters, and the Adam optimizer is used to update the network parameters. After each forward propagation of the neural network, the error is back propagated and the network weight is updated, and the network objective function is converged after repeated multiple updates.

[0081] Further, step S3: using the network parameters trained in step S2 to detect the target and estimate the distance parameter of the data preprocessed in step S1.

[0082] Simulation experiment:

[0083] In the simulation experiments of this invention, the deep learning algorithm proposed in this invention is implemented in the deep learning toolbox of Matlab R2020b software, and compared with the cell average constant false alarm rate detector (CFAR). The target detection performance under different conditions is compared and analyzed.

[0084] Simulation conditions: 100 distance units are generated in the fast time dimension, so the network input size is 10×10×1. There is one target, and the noise is additive white Gaussian noise, with a signal-to-noise ratio (SNR) of 10 dB for the training samples. Each SNR generates 6000 samples, with 6000 samples containing the target and 6000 samples without the target, resulting in a total of 12,000 samples. The training and test sets are generated at a 10:2 ratio. Statistical methods are used for the detection probability and false alarm probability in deep learning. The learning hyperparameters of the convolutional neural network are shown in Table 1 below.

[0085] Table 1. Learning hyperparameters of convolutional neural networks

[0086]

[0087] Simulation results:

[0088] (1) Learning rate curve

[0089] The learning rate is a crucial parameter controlling the step size for updating model weights. During training, the learning rate is gradually adjusted as training progresses. The graph showing the learning rate variation during training of the convolutional neural network in this invention is attached. Figure 3 As shown, from Figure 3 The process clearly shows that the learning rate gradually decreases as the training rounds progress.

[0090] (2) Loss function curve

[0091] During training, the loss function gradually decreases as training progresses. The loss values ​​during the training process of the convolutional neural network in this invention are plotted as curves, and the results are shown in the appendix. Figure 4 As shown, from the appendix Figure 4 The decreasing trend of the loss function can be observed intuitively, and the system eventually trains to a state with a good fit.

[0092] (3) Target output results

[0093] The trained network is used to detect targets and estimate their locations on incoming wave data. Figure 5 and attached Figure 6 The network outputs with and without a target are presented respectively. (From the appendix...) Figure 5 and attached Figure 6It can be seen that, under the condition of 10dB, the output probability of no target is less than 0.5, and no false alarm is generated. The output probability of the target conforms to the target distance estimation spectrum diagram, and the target can be successfully detected. The attached Figure 7 The estimation result graph of 50 target estimations is given, and it can be seen that the target value can be correctly estimated.

[0094] The detection result of the radar single target detection method in the application is compared with the CA-CFAR detection algorithm. Based on statistical signal theory, the detection threshold value commonly used in engineering is about 13.5 decibels, that is, the threshold factor K of the CA-CFAR detection is 4.5, so 10log10(4.5)

[0095] = 13.06 decibels. Assuming that the signal-to-noise ratio received by the radar is 10 decibels, 1000 detections are performed in the simulation experiment, and the detection probability (Pd) and the false alarm rate (Pfa) are counted. The comparison results of the two methods are shown in Table 2.

[0096] Table 2 Comparison results of the detection method in the application and the CA-CFAR detection method

[0097] CA-CFAR Method in the invention Detection probability (Pd) 63.60％ 91.02％ False alarm rate (Pfa) 7.70％ 3.46％

[0098] As can be seen from Table 2, the detection probability of the detection method in the application for the radar target is much greater than that of the CA-CFAR method, and the false alarm rate is less than that of the CA-CFAR method.

[0099] Embodiment two:

[0100] Embodiment two provides a radar single target detection system based on a convolutional neural network, which comprises a radar data preprocessing module and a convolutional neural network module. The radar data preprocessing module is used for preprocessing radar echo data to obtain a one-dimensional fast time distance sequence of a target, and the convolutional neural network module is used for detecting a radar single target.

[0101] The radar data preprocessing module and the convolutional neural network module are realized based on the detection method described in embodiment one.

[0102] Embodiment three:

[0103] Embodiment three provides an electronic device, which comprises at least one processor, and a memory connected with the processor in communication. The memory stores instructions executable by the processor, and the instructions are executed by the processor to enable the processor to execute the detection method described in embodiment one.

[0104] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A radar single-target detection method based on convolutional neural networks, characterized in that, Includes the following steps: S1: Preprocess the radar echo data to obtain a one-dimensional fast time-range sequence of the target; A one-dimensional fast time-distance sequence is pulse-compressed and then sampled to obtain the detection signal. One distance unit; S2: Construct and train a convolutional neural network for detecting single radar targets; The detected signal The convolutional neural network is input with distance units, and the input data dimension is... That is, length × width is The number of channels is 1, and the output is... A 3D signal vector, with the positions of non-zero elements in the signal vector, i.e., the target distance, as the label, denoted as . The input data and output label pairs of the convolutional neural network are then represented as follows: ； in, This represents the training sample size of the network input. For tags, Let be the number of samples. If a target is detected, then... This is a zero-one vector, and the positions of the non-zero elements represent the target distance. If no target is detected, then... It is a vector of all zeros; S3: Use the network parameters trained in step S2 to perform target detection and distance parameter estimation on the preprocessed data in step S1.

2. The radar single-target detection method based on convolutional neural networks according to claim 1, characterized in that, Step S1 includes the following steps: S101: Establish a model of the return signal after the radar transmitted signal encounters a target; S102: The radar's return signal is down-converted and then subjected to orthogonal phase detection to obtain the echo signal model, which is the target's one-dimensional fast time-range sequence.

3. The radar single-target detection method based on convolutional neural networks according to claim 2, characterized in that, The echo signal model described in step S102 is represented as follows: ； in, For echo signal, For noise; A is the amplitude of the echo; This is the delay of the echo signal; The delay caused by the target, R t Let c be the distance to the target and c be the speed of light. The target Doppler frequency, For the target radial velocity, This refers to the operating wavelength of the radar.

4. The radar single-target detection method based on convolutional neural networks according to claim 3, characterized in that, The convolutional neural network described in step S2 includes six convolutional layers and two fully connected layers. Each convolutional layer has 256 convolutional kernels, and the kernel size of the first two convolutional layers is [missing information]. The kernel size of the last four convolutional layers is Each convolutional layer is followed by a batch normalization layer and a ReLU layer; after all the convolutional layers are two fully connected layers. After the first fully connected layer is a ReLU activation function and a dropput layer, and after the second fully connected layer is a sigmoid activation function to induce classification probabilities and output them.

5. The radar single-target detection method based on convolutional neural networks according to claim 4, characterized in that: The parameters of the convolutional layer network of the convolutional neural network are as follows: ,in The dimension is convolution kernel, Indicates the number of input channels. This represents the length and width of the convolution kernel. This indicates the number of convolution kernels, which is also the number of output channels. The dimension is The bias vector.

6. The radar single-target detection method based on convolutional neural networks according to claim 5, characterized in that: The network parameters of the fully connected layer of the convolutional neural network are: ,in, The dimension is The weight matrix, Indicates the number of input channels. This indicates the number of neurons in the fully connected layer, which is also the number of output channels. The dimension is The bias vector.

7. A radar single-target detection system based on convolutional neural networks, characterized in that: It includes a radar data preprocessing module and a convolutional neural network module. The radar data preprocessing module is used to preprocess radar echo data to obtain a one-dimensional fast time-range sequence of the target. The convolutional neural network module is used to detect single radar targets. The radar data preprocessing module and the convolutional neural network module are implemented based on the detection method described in any one of claims 1-6.

8. An electronic device, characterized in that: It includes at least one processor; and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the detection method according to any one of claims 1-6.

Images